Update README.md

ljspeech/hificaptain from #99

.github/dependabot.yml

@@ -7,6 +7,7 @@ version: 2
 updates:
   - package-ecosystem: "pip" # See documentation for possible values
     directory: "/" # Location of package manifests
+    target-branch: "dev"
     schedule:
       interval: "daily"
     ignore:

.gitignore

@@ -161,3 +161,4 @@ generator_v1
 g_02500000
 gradio_cached_examples/
 synth_output/
+/data

.pre-commit-config.yaml

@@ -1,9 +1,9 @@
 default_language_version:
-  python: python3.10
+  python: python3.11
 
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.4.0
+    rev: v4.5.0
     hooks:
       # list of supported hooks: https://pre-commit.com/hooks.html
       - id: trailing-whitespace
@@ -18,28 +18,28 @@ repos:
 
   # python code formatting
   - repo: https://github.com/psf/black
-    rev: 23.1.0
+    rev: 23.12.1
     hooks:
       - id: black
         args: [--line-length, "120"]
 
   # python import sorting
   - repo: https://github.com/PyCQA/isort
-    rev: 5.12.0
+    rev: 5.13.2
     hooks:
       - id: isort
         args: ["--profile", "black", "--filter-files"]
 
   # python upgrading syntax to newer version
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.3.1
+    rev: v3.15.0
     hooks:
       - id: pyupgrade
         args: [--py38-plus]
 
   # python check (PEP8), programming errors and code complexity
   - repo: https://github.com/PyCQA/flake8
-    rev: 6.0.0
+    rev: 7.0.0
     hooks:
       - id: flake8
         args:
@@ -54,6 +54,6 @@ repos:
 
   # pylint
   - repo: https://github.com/pycqa/pylint
-    rev: v2.8.2
+    rev: v3.0.3
     hooks:
     -   id: pylint

.pylintrc

@@ -82,16 +82,6 @@ disable=missing-docstring,
         no-name-in-module,
         no-member,
         unsubscriptable-object,
-        print-statement,
-        parameter-unpacking,
-        unpacking-in-except,
-        old-raise-syntax,
-        backtick,
-        long-suffix,
-        old-ne-operator,
-        old-octal-literal,
-        import-star-module-level,
-        non-ascii-bytes-literal,
         raw-checker-failed,
         bad-inline-option,
         locally-disabled,
@@ -106,67 +96,6 @@ disable=missing-docstring,
         too-many-arguments,
         too-many-locals,
         too-many-statements,
-        apply-builtin,
-        basestring-builtin,
-        buffer-builtin,
-        cmp-builtin,
-        coerce-builtin,
-        execfile-builtin,
-        file-builtin,
-        long-builtin,
-        raw_input-builtin,
-        reduce-builtin,
-        standarderror-builtin,
-        unicode-builtin,
-        xrange-builtin,
-        coerce-method,
-        delslice-method,
-        getslice-method,
-        setslice-method,
-        no-absolute-import,
-        old-division,
-        dict-iter-method,
-        dict-view-method,
-        next-method-called,
-        metaclass-assignment,
-        indexing-exception,
-        raising-string,
-        reload-builtin,
-        oct-method,
-        hex-method,
-        nonzero-method,
-        cmp-method,
-        input-builtin,
-        round-builtin,
-        intern-builtin,
-        unichr-builtin,
-        map-builtin-not-iterating,
-        zip-builtin-not-iterating,
-        range-builtin-not-iterating,
-        filter-builtin-not-iterating,
-        using-cmp-argument,
-        eq-without-hash,
-        div-method,
-        idiv-method,
-        rdiv-method,
-        exception-message-attribute,
-        invalid-str-codec,
-        sys-max-int,
-        bad-python3-import,
-        deprecated-string-function,
-        deprecated-str-translate-call,
-        deprecated-itertools-function,
-        deprecated-types-field,
-        next-method-defined,
-        dict-items-not-iterating,
-        dict-keys-not-iterating,
-        dict-values-not-iterating,
-        deprecated-operator-function,
-        deprecated-urllib-function,
-        xreadlines-attribute,
-        deprecated-sys-function,
-        exception-escape,
-        comprehension-escape,
         duplicate-code,
         not-callable,
         import-outside-toplevel,
@@ -363,13 +292,6 @@ max-line-length=120
 # Maximum number of lines in a module.
 max-module-lines=1000
 
-# List of optional constructs for which whitespace checking is disabled. `dict-
-# separator` is used to allow tabulation in dicts, etc.: {1  : 1,\n222: 2}.
-# `trailing-comma` allows a space between comma and closing bracket: (a, ).
-# `empty-line` allows space-only lines.
-no-space-check=trailing-comma,
-               dict-separator
-
 # Allow the body of a class to be on the same line as the declaration if body
 # contains single statement.
 single-line-class-stmt=no
@@ -599,5 +521,5 @@ min-public-methods=2
 
 # Exceptions that will emit a warning when being caught. Defaults to
 # "BaseException, Exception".
-overgeneral-exceptions=BaseException,
+overgeneral-exceptions=builtins.BaseException,
-                       Exception
+                       builtins.Exception

Makefile

@@ -17,7 +17,7 @@ create-package: ## Create wheel and tar gz
 	rm -rf dist/
 	python setup.py bdist_wheel --plat-name=manylinux1_x86_64
 	python setup.py sdist
-	python -m twine upload  dist/* --verbose
+	python -m twine upload  dist/* --verbose --skip-existing
 
 format: ## Run pre-commit hooks
 	pre-commit run -a

README.md

@@ -10,29 +10,31 @@
 [![hydra](https://img.shields.io/badge/Config-Hydra_1.3-89b8cd)](https://hydra.cc/)
 [![black](https://img.shields.io/badge/Code%20Style-Black-black.svg?labelColor=gray)](https://black.readthedocs.io/en/stable/)
 [![isort](https://img.shields.io/badge/%20imports-isort-%231674b1?style=flat&labelColor=ef8336)](https://pycqa.github.io/isort/)
-
+[![PyPI Downloads](https://static.pepy.tech/personalized-badge/matcha-tts?period=total&units=INTERNATIONAL_SYSTEM&left_color=BLACK&right_color=GREEN&left_text=downloads)](https://pepy.tech/projects/matcha-tts)
 <p style="text-align: center;">
   <img src="https://shivammehta25.github.io/Matcha-TTS/images/logo.png" height="128"/>
 </p>
 
 </div>
 
-> This is the official code implementation of 🍵 Matcha-TTS.
+> This is the official code implementation of 🍵 Matcha-TTS [ICASSP 2024].
 
-We propose 🍵 Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses conditional flow matching (similar to rectified flows) to speed up ODE-based speech synthesis. Our method:
+We propose 🍵 Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses [conditional flow matching](https://arxiv.org/abs/2210.02747) (similar to [rectified flows](https://arxiv.org/abs/2209.03003)) to speed up ODE-based speech synthesis. Our method:
 
 - Is probabilistic
 - Has compact memory footprint
 - Sounds highly natural
 - Is very fast to synthesise from
 
-Check out our [demo page](https://shivammehta25.github.io/Matcha-TTS). Read our [arXiv preprint for more details](https://arxiv.org/abs/2309.03199).
+Check out our [demo page](https://shivammehta25.github.io/Matcha-TTS) and read [our ICASSP 2024 paper](https://arxiv.org/abs/2309.03199) for more details.
 
-[Pretrained models](https://drive.google.com/drive/folders/17C_gYgEHOxI5ZypcfE_k1piKCtyR0isJ?usp=sharing) will be auto downloaded with the CLI or gradio interface.
+[Pre-trained models](https://drive.google.com/drive/folders/17C_gYgEHOxI5ZypcfE_k1piKCtyR0isJ?usp=sharing) will be automatically downloaded with the CLI or gradio interface.
 
-[Try 🍵 Matcha-TTS on HuggingFace 🤗 spaces!](https://huggingface.co/spaces/shivammehta25/Matcha-TTS)
+You can also [try 🍵 Matcha-TTS in your browser on HuggingFace 🤗 spaces](https://huggingface.co/spaces/shivammehta25/Matcha-TTS).
 
-<br>
+## Teaser video
+
+[![Watch the video](https://img.youtube.com/vi/xmvJkz3bqw0/hqdefault.jpg)](https://youtu.be/xmvJkz3bqw0)
 
 ## Installation
 
@@ -43,7 +45,7 @@ conda create -n matcha-tts python=3.10 -y
 conda activate matcha-tts
 ```
 
-2. Install Matcha TTS using pip  or from source
+2. Install Matcha TTS using pip or from source
 
 ```bash
 pip install matcha-tts
@@ -53,6 +55,8 @@ from source
 
 ```bash
 pip install git+https://github.com/shivammehta25/Matcha-TTS.git
+cd Matcha-TTS
+pip install -e .
 ```
 
 3. Run CLI / gradio app / jupyter notebook
@@ -110,26 +114,13 @@ matcha-tts --text "<INPUT TEXT>" --temperature 0.667
 matcha-tts --text "<INPUT TEXT>" --steps 10
 ```
 
-## Citation information
-
-If you find this work useful, please cite our paper:
-
-```text
-@article{mehta2023matcha,
-  title={Matcha-TTS: A fast TTS architecture with conditional flow matching},
-  author={Mehta, Shivam and Tu, Ruibo and Beskow, Jonas and Sz{\'e}kely, {\'E}va and Henter, Gustav Eje},
-  journal={arXiv preprint arXiv:2309.03199},
-  year={2023}
-}
-```
-
 ## Train with your own dataset
 
-Let's assume we are training with LJSpeech
+Let's assume we are training with LJ Speech
 
-1. Download the dataset from [here](https://keithito.com/LJ-Speech-Dataset/), extract it to `data/LJSpeech-1.1`, and prepare the filelists to point to the extracted data like the [5th point of setup in Tacotron2 repo](https://github.com/NVIDIA/tacotron2#setup).
+1. Download the dataset from [here](https://keithito.com/LJ-Speech-Dataset/), extract it to `data/LJSpeech-1.1`, and prepare the file lists to point to the extracted data like for [item 5 in the setup of the NVIDIA Tacotron 2 repo](https://github.com/NVIDIA/tacotron2#setup).
 
-2. Clone and enter this repository
+2. Clone and enter the Matcha-TTS repository
 
 ```bash
 git clone https://github.com/shivammehta25/Matcha-TTS.git
@@ -167,7 +158,7 @@ data_statistics:  # Computed for ljspeech dataset
 
 to the paths of your train and validation filelists.
 
-5. Run the training script
+6. Run the training script
 
 ```bash
 make train-ljspeech
@@ -191,20 +182,134 @@ python matcha/train.py experiment=ljspeech_min_memory
 python matcha/train.py experiment=ljspeech trainer.devices=[0,1]
 ```
 
-6. Synthesise from the custom trained model
+7. Synthesise from the custom trained model
 
 ```bash
 matcha-tts --text "<INPUT TEXT>" --checkpoint_path <PATH TO CHECKPOINT>
 ```
 
+## ONNX support
+
+> Special thanks to [@mush42](https://github.com/mush42) for implementing ONNX export and inference support.
+
+It is possible to export Matcha checkpoints to [ONNX](https://onnx.ai/), and run inference on the exported ONNX graph.
+
+### ONNX export
+
+To export a checkpoint to ONNX, first install ONNX with
+
+```bash
+pip install onnx
+```
+
+then run the following:
+
+```bash
+python3 -m matcha.onnx.export matcha.ckpt model.onnx --n-timesteps 5
+```
+
+Optionally, the ONNX exporter accepts **vocoder-name** and **vocoder-checkpoint** arguments. This enables you to embed the vocoder in the exported graph and generate waveforms in a single run (similar to end-to-end TTS systems).
+
+**Note** that `n_timesteps` is treated as a hyper-parameter rather than a model input. This means you should specify it during export (not during inference). If not specified, `n_timesteps` is set to **5**.
+
+**Important**: for now, torch>=2.1.0 is needed for export since the `scaled_product_attention` operator is not exportable in older versions. Until the final version is released, those who want to export their models must install torch>=2.1.0 manually as a pre-release.
+
+### ONNX Inference
+
+To run inference on the exported model, first install `onnxruntime` using
+
+```bash
+pip install onnxruntime
+pip install onnxruntime-gpu  # for GPU inference
+```
+
+then use the following:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs
+```
+
+You can also control synthesis parameters:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --temperature 0.4 --speaking_rate 0.9 --spk 0
+```
+
+To run inference on **GPU**, make sure to install **onnxruntime-gpu** package, and then pass `--gpu` to the inference command:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --gpu
+```
+
+If you exported only Matcha to ONNX, this will write mel-spectrogram as graphs and `numpy` arrays to the output directory.
+If you embedded the vocoder in the exported graph, this will write `.wav` audio files to the output directory.
+
+If you exported only Matcha to ONNX, and you want to run a full TTS pipeline, you can pass a path to a vocoder model in `ONNX` format:
+
+```bash
+python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --vocoder hifigan.small.onnx
+```
+
+This will write `.wav` audio files to the output directory.
+
+## Extract phoneme alignments from Matcha-TTS
+
+If the dataset is structured as
+
+```bash
+data/
+└── LJSpeech-1.1
+    ├── metadata.csv
+    ├── README
+    ├── test.txt
+    ├── train.txt
+    ├── val.txt
+    └── wavs
+```
+Then you can extract the phoneme level alignments from a Trained Matcha-TTS model using:
+```bash
+python  matcha/utils/get_durations_from_trained_model.py -i dataset_yaml -c <checkpoint>
+```
+Example:
+```bash
+python  matcha/utils/get_durations_from_trained_model.py -i ljspeech.yaml -c matcha_ljspeech.ckpt
+```
+or simply:
+```bash
+matcha-tts-get-durations -i ljspeech.yaml -c matcha_ljspeech.ckpt
+```
+---
+## Train using extracted alignments
+
+In the datasetconfig turn on load duration.
+Example: `ljspeech.yaml`
+```
+load_durations: True
+```
+or see an examples in configs/experiment/ljspeech_from_durations.yaml
+
+
+## Citation information
+
+If you use our code or otherwise find this work useful, please cite our paper:
+
+```text
+@inproceedings{mehta2024matcha,
+  title={Matcha-{TTS}: A fast {TTS} architecture with conditional flow matching},
+  author={Mehta, Shivam and Tu, Ruibo and Beskow, Jonas and Sz{\'e}kely, {\'E}va and Henter, Gustav Eje},
+  booktitle={Proc. ICASSP},
+  year={2024}
+}
+```
+
 ## Acknowledgements
 
-Since this code uses: [Lightning-Hydra-Template](https://github.com/ashleve/lightning-hydra-template), you have all the powers that comes with it.
+Since this code uses [Lightning-Hydra-Template](https://github.com/ashleve/lightning-hydra-template), you have all the powers that come with it.
 
-Other source codes I would like to acknowledge:
+Other source code we would like to acknowledge:
 
-- [Coqui-TTS](https://github.com/coqui-ai/TTS/tree/dev) :For helping me figure out how to make cython binaries pip installable and encouragement
+- [Coqui-TTS](https://github.com/coqui-ai/TTS/tree/dev): For helping me figure out how to make cython binaries pip installable and encouragement
 - [Hugging Face Diffusers](https://huggingface.co/): For their awesome diffusers library and its components
-- [Grad-TTS](https://github.com/huawei-noah/Speech-Backbones/tree/main/Grad-TTS): For source code of MAS
+- [Grad-TTS](https://github.com/huawei-noah/Speech-Backbones/tree/main/Grad-TTS): For the monotonic alignment search source code
 - [torchdyn](https://github.com/DiffEqML/torchdyn): Useful for trying other ODE solvers during research and development
-- [labml.ai](https://nn.labml.ai/transformers/rope/index.html): For RoPE implementation
+- [labml.ai](https://nn.labml.ai/transformers/rope/index.html): For the RoPE implementation

configs/data/hi-fi_en-US_female.yaml

@@ -0,0 +1,14 @@
+defaults:
+  - ljspeech
+  - _self_
+
+# Dataset URL: https://ast-astrec.nict.go.jp/en/release/hi-fi-captain/
+_target_: matcha.data.text_mel_datamodule.TextMelDataModule
+name: hi-fi_en-US_female
+train_filelist_path: data/hi-fi_en-US_female/train.txt
+valid_filelist_path: data/hi-fi_en-US_female/val.txt
+batch_size: 32
+cleaners: [english_cleaners_piper]
+data_statistics:  # Computed for this dataset
+  mel_mean: -6.38385
+  mel_std: 2.541796

configs/data/ljspeech.yaml

@@ -1,7 +1,7 @@
 _target_: matcha.data.text_mel_datamodule.TextMelDataModule
 name: ljspeech
-train_filelist_path: data/filelists/ljs_audio_text_train_filelist.txt
+train_filelist_path: data/LJSpeech-1.1/train.txt
-valid_filelist_path: data/filelists/ljs_audio_text_val_filelist.txt
+valid_filelist_path: data/LJSpeech-1.1/val.txt
 batch_size: 32
 num_workers: 20
 pin_memory: True
@@ -19,3 +19,4 @@ data_statistics:  # Computed for ljspeech dataset
   mel_mean: -5.536622
   mel_std: 2.116101
 seed: ${seed}
+load_durations: false

configs/debug/default.yaml

@@ -7,8 +7,8 @@
 task_name: "debug"
 
 # disable callbacks and loggers during debugging
-callbacks: null
+# callbacks: null
-logger: null
+# logger: null
 
 extras:
   ignore_warnings: False

configs/debug/profiler.yaml

@@ -7,6 +7,9 @@ defaults:
 
 trainer:
   max_epochs: 1
-  profiler: "simple"
+  # profiler: "simple"
-  # profiler: "advanced"
+  profiler: "advanced"
   # profiler: "pytorch"
+  accelerator: gpu
+
+  limit_train_batches: 0.02

configs/experiment/hifi_dataset_piper_phonemizer.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=multispeaker
+
+defaults:
+  - override /data: hi-fi_en-US_female.yaml
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["hi-fi", "single_speaker", "piper_phonemizer", "en_US", "female"]
+
+run_name: hi-fi_en-US_female_piper_phonemizer

configs/experiment/ljspeech_from_durations.yaml

@@ -0,0 +1,19 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=multispeaker
+
+defaults:
+  - override /data: ljspeech.yaml
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["ljspeech"]
+
+run_name: ljspeech
+
+
+data:
+  load_durations: True
+  batch_size: 64

configs/model/matcha.yaml

@@ -12,3 +12,5 @@ spk_emb_dim: 64
 n_feats: 80
 data_statistics: ${data.data_statistics}
 out_size: null # Must be divisible by 4
+prior_loss: true
+use_precomputed_durations: ${data.load_durations}

data

@@ -1 +0,0 @@
-/home/smehta/Projects/Speech-Backbones/Grad-TTS/data

matcha/VERSION

@@ -1 +1 @@
-0.0.1.dev4
+0.0.7.2

matcha/app.py

@@ -8,7 +8,7 @@ import torch
 
 from matcha.cli import (
     MATCHA_URLS,
-    VOCODER_URL,
+    VOCODER_URLS,
     assert_model_downloaded,
     get_device,
     load_matcha,
@@ -22,20 +22,80 @@ LOCATION = Path(get_user_data_dir())
 
 args = Namespace(
     cpu=False,
-    model="matcha_ljspeech",
+    model="matcha_vctk",
-    vocoder="hifigan_T2_v1",
+    vocoder="hifigan_univ_v1",
-    spk=None,
+    spk=0,
 )
 
-MATCHA_TTS_LOC = LOCATION / f"{args.model}.ckpt"
+CURRENTLY_LOADED_MODEL = args.model
-VOCODER_LOC = LOCATION / f"{args.vocoder}"
+
+
+def MATCHA_TTS_LOC(x):
+    return LOCATION / f"{x}.ckpt"
+
+
+def VOCODER_LOC(x):
+    return LOCATION / f"{x}"
+
+
 LOGO_URL = "https://shivammehta25.github.io/Matcha-TTS/images/logo.png"
-assert_model_downloaded(MATCHA_TTS_LOC, MATCHA_URLS[args.model])
+RADIO_OPTIONS = {
-assert_model_downloaded(VOCODER_LOC, VOCODER_URL[args.vocoder])
+    "Multi Speaker (VCTK)": {
+        "model": "matcha_vctk",
+        "vocoder": "hifigan_univ_v1",
+    },
+    "Single Speaker (LJ Speech)": {
+        "model": "matcha_ljspeech",
+        "vocoder": "hifigan_T2_v1",
+    },
+}
+
+# Ensure all the required models are downloaded
+assert_model_downloaded(MATCHA_TTS_LOC("matcha_ljspeech"), MATCHA_URLS["matcha_ljspeech"])
+assert_model_downloaded(VOCODER_LOC("hifigan_T2_v1"), VOCODER_URLS["hifigan_T2_v1"])
+assert_model_downloaded(MATCHA_TTS_LOC("matcha_vctk"), MATCHA_URLS["matcha_vctk"])
+assert_model_downloaded(VOCODER_LOC("hifigan_univ_v1"), VOCODER_URLS["hifigan_univ_v1"])
+
 device = get_device(args)
 
-model = load_matcha(args.model, MATCHA_TTS_LOC, device)
+# Load default model
-vocoder, denoiser = load_vocoder(args.vocoder, VOCODER_LOC, device)
+model = load_matcha(args.model, MATCHA_TTS_LOC(args.model), device)
+vocoder, denoiser = load_vocoder(args.vocoder, VOCODER_LOC(args.vocoder), device)
+
+
+def load_model(model_name, vocoder_name):
+    model = load_matcha(model_name, MATCHA_TTS_LOC(model_name), device)
+    vocoder, denoiser = load_vocoder(vocoder_name, VOCODER_LOC(vocoder_name), device)
+    return model, vocoder, denoiser
+
+
+def load_model_ui(model_type, textbox):
+    model_name, vocoder_name = RADIO_OPTIONS[model_type]["model"], RADIO_OPTIONS[model_type]["vocoder"]
+
+    global model, vocoder, denoiser, CURRENTLY_LOADED_MODEL  # pylint: disable=global-statement
+    if CURRENTLY_LOADED_MODEL != model_name:
+        model, vocoder, denoiser = load_model(model_name, vocoder_name)
+        CURRENTLY_LOADED_MODEL = model_name
+
+    if model_name == "matcha_ljspeech":
+        spk_slider = gr.update(visible=False, value=-1)
+        single_speaker_examples = gr.update(visible=True)
+        multi_speaker_examples = gr.update(visible=False)
+        length_scale = gr.update(value=0.95)
+    else:
+        spk_slider = gr.update(visible=True, value=0)
+        single_speaker_examples = gr.update(visible=False)
+        multi_speaker_examples = gr.update(visible=True)
+        length_scale = gr.update(value=0.85)
+
+    return (
+        textbox,
+        gr.update(interactive=True),
+        spk_slider,
+        single_speaker_examples,
+        multi_speaker_examples,
+        length_scale,
+    )
 
 
 @torch.inference_mode()
@@ -45,13 +105,14 @@ def process_text_gradio(text):
 
 
 @torch.inference_mode()
-def synthesise_mel(text, text_length, n_timesteps, temperature, length_scale):
+def synthesise_mel(text, text_length, n_timesteps, temperature, length_scale, spk):
+    spk = torch.tensor([spk], device=device, dtype=torch.long) if spk >= 0 else None
     output = model.synthesise(
         text,
         text_length,
         n_timesteps=n_timesteps,
         temperature=temperature,
-        spks=args.spk,
+        spks=spk,
         length_scale=length_scale,
     )
     output["waveform"] = to_waveform(output["mel"], vocoder, denoiser)
@@ -61,9 +122,27 @@ def synthesise_mel(text, text_length, n_timesteps, temperature, length_scale):
     return fp.name, plot_tensor(output["mel"].squeeze().cpu().numpy())
 
 
-def run_full_synthesis(text, n_timesteps, mel_temp, length_scale):
+def multispeaker_example_cacher(text, n_timesteps, mel_temp, length_scale, spk):
+    global CURRENTLY_LOADED_MODEL  # pylint: disable=global-statement
+    if CURRENTLY_LOADED_MODEL != "matcha_vctk":
+        global model, vocoder, denoiser  # pylint: disable=global-statement
+        model, vocoder, denoiser = load_model("matcha_vctk", "hifigan_univ_v1")
+        CURRENTLY_LOADED_MODEL = "matcha_vctk"
+
     phones, text, text_lengths = process_text_gradio(text)
-    audio, mel_spectrogram = synthesise_mel(text, text_lengths, n_timesteps, mel_temp, length_scale)
+    audio, mel_spectrogram = synthesise_mel(text, text_lengths, n_timesteps, mel_temp, length_scale, spk)
+    return phones, audio, mel_spectrogram
+
+
+def ljspeech_example_cacher(text, n_timesteps, mel_temp, length_scale, spk=-1):
+    global CURRENTLY_LOADED_MODEL  # pylint: disable=global-statement
+    if CURRENTLY_LOADED_MODEL != "matcha_ljspeech":
+        global model, vocoder, denoiser  # pylint: disable=global-statement
+        model, vocoder, denoiser = load_model("matcha_ljspeech", "hifigan_T2_v1")
+        CURRENTLY_LOADED_MODEL = "matcha_ljspeech"
+
+    phones, text, text_lengths = process_text_gradio(text)
+    audio, mel_spectrogram = synthesise_mel(text, text_lengths, n_timesteps, mel_temp, length_scale, spk)
     return phones, audio, mel_spectrogram
 
 
@@ -92,20 +171,31 @@ def main():
         with gr.Box():
             with gr.Row():
                 gr.Markdown(description, scale=3)
-                gr.Image(LOGO_URL, label="Matcha-TTS logo", height=150, width=150, scale=1, show_label=False)
+                with gr.Column():
+                    gr.Image(LOGO_URL, label="Matcha-TTS logo", height=50, width=50, scale=1, show_label=False)
+                    html = '<br><iframe width="560" height="315" src="https://www.youtube.com/embed/xmvJkz3bqw0?si=jN7ILyDsbPwJCGoa" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe>'
+                    gr.HTML(html)
 
         with gr.Box():
+            radio_options = list(RADIO_OPTIONS.keys())
+            model_type = gr.Radio(
+                radio_options, value=radio_options[0], label="Choose a Model", interactive=True, container=False
+            )
+
             with gr.Row():
                 gr.Markdown("# Text Input")
             with gr.Row():
-                text = gr.Textbox(value="", lines=2, label="Text to synthesise")
+                text = gr.Textbox(value="", lines=2, label="Text to synthesise", scale=3)
+                spk_slider = gr.Slider(
+                    minimum=0, maximum=107, step=1, value=args.spk, label="Speaker ID", interactive=True, scale=1
+                )
 
             with gr.Row():
                 gr.Markdown("### Hyper parameters")
             with gr.Row():
                 n_timesteps = gr.Slider(
                     label="Number of ODE steps",
-                    minimum=0,
+                    minimum=1,
                     maximum=100,
                     step=1,
                     value=10,
@@ -142,58 +232,110 @@ def main():
                 # with gr.Row():
                 audio = gr.Audio(interactive=False, label="Audio")
 
-        with gr.Row():
+        with gr.Row(visible=False) as example_row_lj_speech:
             examples = gr.Examples(  # pylint: disable=unused-variable
                 examples=[
                     [
                         "We propose Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses conditional flow matching (similar to rectified flows) to speed up O D E-based speech synthesis.",
                         50,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                     [
                         "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
                         2,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                     [
                         "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
                         4,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                     [
                         "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
                         10,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                     [
                         "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
                         50,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                     [
                         "The narrative of these events is based largely on the recollections of the participants.",
                         10,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                     [
                         "The jury did not believe him, and the verdict was for the defendants.",
                         10,
                         0.677,
-                        1.0,
+                        0.95,
                     ],
                 ],
-                fn=run_full_synthesis,
+                fn=ljspeech_example_cacher,
                 inputs=[text, n_timesteps, mel_temp, length_scale],
                 outputs=[phonetised_text, audio, mel_spectrogram],
                 cache_examples=True,
             )
 
+        with gr.Row() as example_row_multispeaker:
+            multi_speaker_examples = gr.Examples(  # pylint: disable=unused-variable
+                examples=[
+                    [
+                        "Hello everyone! I am speaker 0 and I am here to tell you that Matcha-TTS is amazing!",
+                        10,
+                        0.677,
+                        0.85,
+                        0,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 16 and I am here to tell you that Matcha-TTS is amazing!",
+                        10,
+                        0.677,
+                        0.85,
+                        16,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 44 and I am here to tell you that Matcha-TTS is amazing!",
+                        50,
+                        0.677,
+                        0.85,
+                        44,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 45 and I am here to tell you that Matcha-TTS is amazing!",
+                        50,
+                        0.677,
+                        0.85,
+                        45,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 58 and I am here to tell you that Matcha-TTS is amazing!",
+                        4,
+                        0.677,
+                        0.85,
+                        58,
+                    ],
+                ],
+                fn=multispeaker_example_cacher,
+                inputs=[text, n_timesteps, mel_temp, length_scale, spk_slider],
+                outputs=[phonetised_text, audio, mel_spectrogram],
+                cache_examples=True,
+                label="Multi Speaker Examples",
+            )
+
+        model_type.change(lambda x: gr.update(interactive=False), inputs=[synth_btn], outputs=[synth_btn]).then(
+            load_model_ui,
+            inputs=[model_type, text],
+            outputs=[text, synth_btn, spk_slider, example_row_lj_speech, example_row_multispeaker, length_scale],
+        )
+
         synth_btn.click(
             fn=process_text_gradio,
             inputs=[
@@ -204,11 +346,11 @@ def main():
             queue=True,
         ).then(
             fn=synthesise_mel,
-            inputs=[processed_text, processed_text_len, n_timesteps, mel_temp, length_scale],
+            inputs=[processed_text, processed_text_len, n_timesteps, mel_temp, length_scale, spk_slider],
             outputs=[audio, mel_spectrogram],
         )
 
-        demo.queue(concurrency_count=5).launch(share=True)
+        demo.queue().launch(share=True)
 
 
 if __name__ == "__main__":

matcha/cli.py

@@ -1,6 +1,7 @@
 import argparse
 import datetime as dt
 import os
+import warnings
 from pathlib import Path
 
 import matplotlib.pyplot as plt
@@ -17,13 +18,20 @@ from matcha.text import sequence_to_text, text_to_sequence
 from matcha.utils.utils import assert_model_downloaded, get_user_data_dir, intersperse
 
 MATCHA_URLS = {
-    "matcha_ljspeech": "https://drive.google.com/file/d/1BBzmMU7k3a_WetDfaFblMoN18GqQeHCg/view?usp=drive_link"
+    "matcha_ljspeech": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_ljspeech.ckpt",
-}  # , "matcha_vctk": ""}  # Coming soon
+    "matcha_vctk": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_vctk.ckpt",
+}
 
-MULTISPEAKER_MODEL = {"matcha_vctk"}
+VOCODER_URLS = {
-SINGLESPEAKER_MODEL = {"matcha_ljspeech"}
+    "hifigan_T2_v1": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/generator_v1",  # Old url: https://drive.google.com/file/d/14NENd4equCBLyyCSke114Mv6YR_j_uFs/view?usp=drive_link
+    "hifigan_univ_v1": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/g_02500000",  # Old url: https://drive.google.com/file/d/1qpgI41wNXFcH-iKq1Y42JlBC9j0je8PW/view?usp=drive_link
+}
 
-VOCODER_URL = {"hifigan_T2_v1": "https://drive.google.com/file/d/14NENd4equCBLyyCSke114Mv6YR_j_uFs/view?usp=drive_link"}
+MULTISPEAKER_MODEL = {
+    "matcha_vctk": {"vocoder": "hifigan_univ_v1", "speaking_rate": 0.85, "spk": 0, "spk_range": (0, 107)}
+}
+
+SINGLESPEAKER_MODEL = {"matcha_ljspeech": {"vocoder": "hifigan_T2_v1", "speaking_rate": 0.95, "spk": None}}
 
 
 def plot_spectrogram_to_numpy(spectrogram, filename):
@@ -40,7 +48,7 @@ def plot_spectrogram_to_numpy(spectrogram, filename):
 def process_text(i: int, text: str, device: torch.device):
     print(f"[{i}] - Input text: {text}")
     x = torch.tensor(
-        intersperse(text_to_sequence(text, ["english_cleaners2"]), 0),
+        intersperse(text_to_sequence(text, ["english_cleaners2"])[0], 0),
         dtype=torch.long,
         device=device,
     )[None]
@@ -55,17 +63,21 @@ def get_texts(args):
     if args.text:
         texts = [args.text]
     else:
-        with open(args.file) as f:
+        with open(args.file, encoding="utf-8") as f:
             texts = f.readlines()
     return texts
 
 
 def assert_required_models_available(args):
     save_dir = get_user_data_dir()
-    model_path = save_dir / f"{args.model}.ckpt"
+    if not hasattr(args, "checkpoint_path") and args.checkpoint_path is None:
+        model_path = args.checkpoint_path
+    else:
+        model_path = save_dir / f"{args.model}.ckpt"
+        assert_model_downloaded(model_path, MATCHA_URLS[args.model])
+
     vocoder_path = save_dir / f"{args.vocoder}"
-    assert_model_downloaded(model_path, MATCHA_URLS[args.model])
+    assert_model_downloaded(vocoder_path, VOCODER_URLS[args.vocoder])
-    assert_model_downloaded(vocoder_path, VOCODER_URL[args.vocoder])
     return {"matcha": model_path, "vocoder": vocoder_path}
 
 
@@ -81,7 +93,7 @@ def load_hifigan(checkpoint_path, device):
 def load_vocoder(vocoder_name, checkpoint_path, device):
     print(f"[!] Loading {vocoder_name}!")
     vocoder = None
-    if vocoder_name == "hifigan_T2_v1":
+    if vocoder_name in ("hifigan_T2_v1", "hifigan_univ_v1"):
         vocoder = load_hifigan(checkpoint_path, device)
     else:
         raise NotImplementedError(
@@ -102,10 +114,10 @@ def load_matcha(model_name, checkpoint_path, device):
     return model
 
 
-def to_waveform(mel, vocoder, denoiser=None):
+def to_waveform(mel, vocoder, denoiser=None, denoiser_strength=0.00025):
     audio = vocoder(mel).clamp(-1, 1)
     if denoiser is not None:
-        audio = denoiser(audio.squeeze(), strength=0.00025).cpu().squeeze()
+        audio = denoiser(audio.squeeze(), strength=denoiser_strength).cpu().squeeze()
 
     return audio.cpu().squeeze()
 
@@ -124,21 +136,70 @@ def validate_args(args):
         args.text or args.file
     ), "Either text or file must be provided Matcha-T(ea)TTS need sometext to whisk the waveforms."
     assert args.temperature >= 0, "Sampling temperature cannot be negative"
-    assert args.speaking_rate > 0, "Speaking rate must be greater than 0"
     assert args.steps > 0, "Number of ODE steps must be greater than 0"
-    if args.model in SINGLESPEAKER_MODEL:
-        assert args.spk is None, f"Speaker ID is not supported for {args.model}"
-    if args.spk is not None:
-        assert args.spk >= 0 and args.spk < 109, "Speaker ID must be between 0 and 108"
-        assert args.model in MULTISPEAKER_MODEL, "Speaker ID is only supported for multispeaker model"
 
-    if args.model in MULTISPEAKER_MODEL:
+    if args.checkpoint_path is None:
-        if args.spk is None:
+        # When using pretrained models
-            print("[!] Speaker ID not provided! Using speaker ID 0")
+        if args.model in SINGLESPEAKER_MODEL:
-            args.spk = 0
+            args = validate_args_for_single_speaker_model(args)
+
+        if args.model in MULTISPEAKER_MODEL:
+            args = validate_args_for_multispeaker_model(args)
+    else:
+        # When using a custom model
+        if args.vocoder != "hifigan_univ_v1":
+            warn_ = "[-] Using custom model checkpoint! I would suggest passing --vocoder hifigan_univ_v1, unless the custom model is trained on LJ Speech."
+            warnings.warn(warn_, UserWarning)
+        if args.speaking_rate is None:
+            args.speaking_rate = 1.0
 
     if args.batched:
         assert args.batch_size > 0, "Batch size must be greater than 0"
+    assert args.speaking_rate > 0, "Speaking rate must be greater than 0"
+
+    return args
+
+
+def validate_args_for_multispeaker_model(args):
+    if args.vocoder is not None:
+        if args.vocoder != MULTISPEAKER_MODEL[args.model]["vocoder"]:
+            warn_ = f"[-] Using {args.model} model! I would suggest passing --vocoder {MULTISPEAKER_MODEL[args.model]['vocoder']}"
+            warnings.warn(warn_, UserWarning)
+    else:
+        args.vocoder = MULTISPEAKER_MODEL[args.model]["vocoder"]
+
+    if args.speaking_rate is None:
+        args.speaking_rate = MULTISPEAKER_MODEL[args.model]["speaking_rate"]
+
+    spk_range = MULTISPEAKER_MODEL[args.model]["spk_range"]
+    if args.spk is not None:
+        assert (
+            args.spk >= spk_range[0] and args.spk <= spk_range[-1]
+        ), f"Speaker ID must be between {spk_range} for this model."
+    else:
+        available_spk_id = MULTISPEAKER_MODEL[args.model]["spk"]
+        warn_ = f"[!] Speaker ID not provided! Using speaker ID {available_spk_id}"
+        warnings.warn(warn_, UserWarning)
+        args.spk = available_spk_id
+
+    return args
+
+
+def validate_args_for_single_speaker_model(args):
+    if args.vocoder is not None:
+        if args.vocoder != SINGLESPEAKER_MODEL[args.model]["vocoder"]:
+            warn_ = f"[-] Using {args.model} model! I would suggest passing --vocoder {SINGLESPEAKER_MODEL[args.model]['vocoder']}"
+            warnings.warn(warn_, UserWarning)
+    else:
+        args.vocoder = SINGLESPEAKER_MODEL[args.model]["vocoder"]
+
+    if args.speaking_rate is None:
+        args.speaking_rate = SINGLESPEAKER_MODEL[args.model]["speaking_rate"]
+
+    if args.spk != SINGLESPEAKER_MODEL[args.model]["spk"]:
+        warn_ = f"[-] Ignoring speaker id {args.spk} for {args.model}"
+        warnings.warn(warn_, UserWarning)
+        args.spk = SINGLESPEAKER_MODEL[args.model]["spk"]
 
     return args
 
@@ -166,9 +227,9 @@ def cli():
     parser.add_argument(
         "--vocoder",
         type=str,
-        default="hifigan_T2_v1",
+        default=None,
-        help="Vocoder to use",
+        help="Vocoder to use (default: will use the one suggested with the pretrained model))",
-        choices=VOCODER_URL.keys(),
+        choices=VOCODER_URLS.keys(),
     )
     parser.add_argument("--text", type=str, default=None, help="Text to synthesize")
     parser.add_argument("--file", type=str, default=None, help="Text file to synthesize")
@@ -182,7 +243,7 @@ def cli():
     parser.add_argument(
         "--speaking_rate",
         type=float,
-        default=1.0,
+        default=None,
         help="change the speaking rate, a higher value means slower speaking rate (default: 1.0)",
     )
     parser.add_argument("--steps", type=int, default=10, help="Number of ODE steps  (default: 10)")
@@ -199,8 +260,10 @@ def cli():
         default=os.getcwd(),
         help="Output folder to save results (default: current dir)",
     )
-    parser.add_argument("--batched", action="store_true")
+    parser.add_argument("--batched", action="store_true", help="Batched inference (default: False)")
-    parser.add_argument("--batch_size", type=int, default=32)
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Batch size only useful when --batched (default: 32)"
+    )
 
     args = parser.parse_args()
 
@@ -263,16 +326,17 @@ def batched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
     for i, batch in enumerate(dataloader):
         i = i + 1
         start_t = dt.datetime.now()
+        b = batch["x"].shape[0]
         output = model.synthesise(
             batch["x"].to(device),
             batch["x_lengths"].to(device),
             n_timesteps=args.steps,
             temperature=args.temperature,
-            spks=spk,
+            spks=spk.expand(b) if spk is not None else spk,
             length_scale=args.speaking_rate,
         )
 
-        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser)
+        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser, args.denoiser_strength)
         t = (dt.datetime.now() - start_t).total_seconds()
         rtf_w = t * 22050 / (output["waveform"].shape[-1])
         print(f"[🍵-Batch: {i}] Matcha-TTS RTF: {output['rtf']:.4f}")
@@ -313,7 +377,7 @@ def unbatched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
             spks=spk,
             length_scale=args.speaking_rate,
         )
-        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser)
+        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser, args.denoiser_strength)
         # RTF with HiFiGAN
         t = (dt.datetime.now() - start_t).total_seconds()
         rtf_w = t * 22050 / (output["waveform"].shape[-1])
@@ -333,6 +397,8 @@ def unbatched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
 
 def print_config(args):
     print("[!] Configurations: ")
+    print(f"\t- Model: {args.model}")
+    print(f"\t- Vocoder: {args.vocoder}")
     print(f"\t- Temperature: {args.temperature}")
     print(f"\t- Speaking rate: {args.speaking_rate}")
     print(f"\t- Number of ODE steps: {args.steps}")

matcha/data/text_mel_datamodule.py

@@ -1,6 +1,8 @@
 import random
+from pathlib import Path
 from typing import Any, Dict, Optional
 
+import numpy as np
 import torch
 import torchaudio as ta
 from lightning import LightningDataModule
@@ -39,6 +41,7 @@ class TextMelDataModule(LightningDataModule):
         f_max,
         data_statistics,
         seed,
+        load_durations,
     ):
         super().__init__()
 
@@ -68,6 +71,7 @@ class TextMelDataModule(LightningDataModule):
             self.hparams.f_max,
             self.hparams.data_statistics,
             self.hparams.seed,
+            self.hparams.load_durations,
         )
         self.validset = TextMelDataset(  # pylint: disable=attribute-defined-outside-init
             self.hparams.valid_filelist_path,
@@ -83,6 +87,7 @@ class TextMelDataModule(LightningDataModule):
             self.hparams.f_max,
             self.hparams.data_statistics,
             self.hparams.seed,
+            self.hparams.load_durations,
         )
 
     def train_dataloader(self):
@@ -109,7 +114,7 @@ class TextMelDataModule(LightningDataModule):
         """Clean up after fit or test."""
         pass  # pylint: disable=unnecessary-pass
 
-    def state_dict(self):  # pylint: disable=no-self-use
+    def state_dict(self):
         """Extra things to save to checkpoint."""
         return {}
 
@@ -134,6 +139,7 @@ class TextMelDataset(torch.utils.data.Dataset):
         f_max=8000,
         data_parameters=None,
         seed=None,
+        load_durations=False,
     ):
         self.filepaths_and_text = parse_filelist(filelist_path)
         self.n_spks = n_spks
@@ -146,6 +152,8 @@ class TextMelDataset(torch.utils.data.Dataset):
         self.win_length = win_length
         self.f_min = f_min
         self.f_max = f_max
+        self.load_durations = load_durations
+
         if data_parameters is not None:
             self.data_parameters = data_parameters
         else:
@@ -164,10 +172,29 @@ class TextMelDataset(torch.utils.data.Dataset):
             filepath, text = filepath_and_text[0], filepath_and_text[1]
             spk = None
 
-        text = self.get_text(text, add_blank=self.add_blank)
+        text, cleaned_text = self.get_text(text, add_blank=self.add_blank)
         mel = self.get_mel(filepath)
 
-        return {"x": text, "y": mel, "spk": spk}
+        durations = self.get_durations(filepath, text) if self.load_durations else None
+
+        return {"x": text, "y": mel, "spk": spk, "filepath": filepath, "x_text": cleaned_text, "durations": durations}
+
+    def get_durations(self, filepath, text):
+        filepath = Path(filepath)
+        data_dir, name = filepath.parent.parent, filepath.stem
+
+        try:
+            dur_loc = data_dir / "durations" / f"{name}.npy"
+            durs = torch.from_numpy(np.load(dur_loc).astype(int))
+
+        except FileNotFoundError as e:
+            raise FileNotFoundError(
+                f"Tried loading the durations but durations didn't exist at {dur_loc}, make sure you've generate the durations first using: python matcha/utils/get_durations_from_trained_model.py \n"
+            ) from e
+
+        assert len(durs) == len(text), f"Length of durations {len(durs)} and text {len(text)} do not match"
+
+        return durs
 
     def get_mel(self, filepath):
         audio, sr = ta.load(filepath)
@@ -187,11 +214,11 @@ class TextMelDataset(torch.utils.data.Dataset):
         return mel
 
     def get_text(self, text, add_blank=True):
-        text_norm = text_to_sequence(text, self.cleaners)
+        text_norm, cleaned_text = text_to_sequence(text, self.cleaners)
         if self.add_blank:
             text_norm = intersperse(text_norm, 0)
         text_norm = torch.IntTensor(text_norm)
-        return text_norm
+        return text_norm, cleaned_text
 
     def __getitem__(self, index):
         datapoint = self.get_datapoint(self.filepaths_and_text[index])
@@ -207,15 +234,18 @@ class TextMelBatchCollate:
 
     def __call__(self, batch):
         B = len(batch)
-        y_max_length = max([item["y"].shape[-1] for item in batch])
+        y_max_length = max([item["y"].shape[-1] for item in batch])  # pylint: disable=consider-using-generator
         y_max_length = fix_len_compatibility(y_max_length)
-        x_max_length = max([item["x"].shape[-1] for item in batch])
+        x_max_length = max([item["x"].shape[-1] for item in batch])  # pylint: disable=consider-using-generator
         n_feats = batch[0]["y"].shape[-2]
 
         y = torch.zeros((B, n_feats, y_max_length), dtype=torch.float32)
         x = torch.zeros((B, x_max_length), dtype=torch.long)
+        durations = torch.zeros((B, x_max_length), dtype=torch.long)
+
         y_lengths, x_lengths = [], []
         spks = []
+        filepaths, x_texts = [], []
         for i, item in enumerate(batch):
             y_, x_ = item["y"], item["x"]
             y_lengths.append(y_.shape[-1])
@@ -223,9 +253,22 @@ class TextMelBatchCollate:
             y[i, :, : y_.shape[-1]] = y_
             x[i, : x_.shape[-1]] = x_
             spks.append(item["spk"])
+            filepaths.append(item["filepath"])
+            x_texts.append(item["x_text"])
+            if item["durations"] is not None:
+                durations[i, : item["durations"].shape[-1]] = item["durations"]
 
         y_lengths = torch.tensor(y_lengths, dtype=torch.long)
         x_lengths = torch.tensor(x_lengths, dtype=torch.long)
         spks = torch.tensor(spks, dtype=torch.long) if self.n_spks > 1 else None
 
-        return {"x": x, "x_lengths": x_lengths, "y": y, "y_lengths": y_lengths, "spks": spks}
+        return {
+            "x": x,
+            "x_lengths": x_lengths,
+            "y": y,
+            "y_lengths": y_lengths,
+            "spks": spks,
+            "filepaths": filepaths,
+            "x_texts": x_texts,
+            "durations": durations if not torch.eq(durations, 0).all() else None,
+        }

matcha/hifigan/denoiser.py

@@ -4,6 +4,10 @@
 import torch
 
 
+class ModeException(Exception):
+    pass
+
+
 class Denoiser(torch.nn.Module):
     """Removes model bias from audio produced with waveglow"""
 
@@ -20,7 +24,7 @@ class Denoiser(torch.nn.Module):
         elif mode == "normal":
             mel_input = torch.randn((1, 80, 88), dtype=dtype, device=device)
         else:
-            raise Exception(f"Mode {mode} if not supported")
+            raise ModeException(f"Mode {mode} if not supported")
 
         def stft_fn(audio, n_fft, hop_length, win_length, window):
             spec = torch.stft(

matcha/hifigan/meldataset.py

@@ -55,7 +55,7 @@ def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin,
     if torch.max(y) > 1.0:
         print("max value is ", torch.max(y))
 
-    global mel_basis, hann_window  # pylint: disable=global-statement
+    global mel_basis, hann_window  # pylint: disable=global-statement,global-variable-not-assigned
     if fmax not in mel_basis:
         mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
         mel_basis[str(fmax) + "_" + str(y.device)] = torch.from_numpy(mel).float().to(y.device)

matcha/hifigan/models.py

@@ -1,7 +1,7 @@
 """ from https://github.com/jik876/hifi-gan """
 
 import torch
-import torch.nn as nn
+import torch.nn as nn  # pylint: disable=consider-using-from-import
 import torch.nn.functional as F
 from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
 from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm

matcha/models/baselightningmodule.py

@@ -58,13 +58,14 @@ class BaseLightningClass(LightningModule, ABC):
         y, y_lengths = batch["y"], batch["y_lengths"]
         spks = batch["spks"]
 
-        dur_loss, prior_loss, diff_loss = self(
+        dur_loss, prior_loss, diff_loss, *_ = self(
             x=x,
             x_lengths=x_lengths,
             y=y,
             y_lengths=y_lengths,
             spks=spks,
             out_size=self.out_size,
+            durations=batch["durations"],
         )
         return {
             "dur_loss": dur_loss,
@@ -81,7 +82,7 @@ class BaseLightningClass(LightningModule, ABC):
             "step",
             float(self.global_step),
             on_step=True,
-            on_epoch=True,
+            prog_bar=True,
             logger=True,
             sync_dist=True,
         )

matcha/models/components/decoder.py

@@ -2,7 +2,7 @@ import math
 from typing import Optional
 
 import torch
-import torch.nn as nn
+import torch.nn as nn  # pylint: disable=consider-using-from-import
 import torch.nn.functional as F
 from conformer import ConformerBlock
 from diffusers.models.activations import get_activation

matcha/models/components/flow_matching.py

@@ -73,16 +73,14 @@ class BASECFM(torch.nn.Module, ABC):
         # Or in future might add like a return_all_steps flag
         sol = []
 
-        steps = 1
+        for step in range(1, len(t_span)):
-        while steps <= len(t_span) - 1:
             dphi_dt = self.estimator(x, mask, mu, t, spks, cond)
 
             x = x + dt * dphi_dt
             t = t + dt
             sol.append(x)
-            if steps < len(t_span) - 1:
+            if step < len(t_span) - 1:
-                dt = t_span[steps + 1] - t
+                dt = t_span[step + 1] - t
-            steps += 1
 
         return sol[-1]
 

matcha/models/components/text_encoder.py

@@ -3,10 +3,10 @@
 import math
 
 import torch
-import torch.nn as nn
+import torch.nn as nn  # pylint: disable=consider-using-from-import
 from einops import rearrange
 
-import matcha.utils as utils
+import matcha.utils as utils  # pylint: disable=consider-using-from-import
 from matcha.utils.model import sequence_mask
 
 log = utils.get_pylogger(__name__)

matcha/models/components/transformer.py

@@ -1,7 +1,7 @@
 from typing import Any, Dict, Optional
 
 import torch
-import torch.nn as nn
+import torch.nn as nn  # pylint: disable=consider-using-from-import
 from diffusers.models.attention import (
     GEGLU,
     GELU,

matcha/models/matcha_tts.py

@@ -4,7 +4,7 @@ import random
 
 import torch
 
-import matcha.utils.monotonic_align as monotonic_align
+import matcha.utils.monotonic_align as monotonic_align  # pylint: disable=consider-using-from-import
 from matcha import utils
 from matcha.models.baselightningmodule import BaseLightningClass
 from matcha.models.components.flow_matching import CFM
@@ -34,6 +34,8 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         out_size,
         optimizer=None,
         scheduler=None,
+        prior_loss=True,
+        use_precomputed_durations=False,
     ):
         super().__init__()
 
@@ -44,6 +46,8 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         self.spk_emb_dim = spk_emb_dim
         self.n_feats = n_feats
         self.out_size = out_size
+        self.prior_loss = prior_loss
+        self.use_precomputed_durations = use_precomputed_durations
 
         if n_spks > 1:
             self.spk_emb = torch.nn.Embedding(n_spks, spk_emb_dim)
@@ -102,6 +106,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
                 # Lengths of mel spectrograms
                 "rtf": float,
                 # Real-time factor
+            }
         """
         # For RTF computation
         t = dt.datetime.now()
@@ -116,7 +121,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         w = torch.exp(logw) * x_mask
         w_ceil = torch.ceil(w) * length_scale
         y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
-        y_max_length = int(y_lengths.max())
+        y_max_length = y_lengths.max()
         y_max_length_ = fix_len_compatibility(y_max_length)
 
         # Using obtained durations `w` construct alignment map `attn`
@@ -145,10 +150,10 @@ class MatchaTTS(BaseLightningClass):  # 🍵
             "rtf": rtf,
         }
 
-    def forward(self, x, x_lengths, y, y_lengths, spks=None, out_size=None, cond=None):
+    def forward(self, x, x_lengths, y, y_lengths, spks=None, out_size=None, cond=None, durations=None):
         """
         Computes 3 losses:
-            1. duration loss: loss between predicted token durations and those extracted by Monotinic Alignment Search (MAS).
+            1. duration loss: loss between predicted token durations and those extracted by Monotonic Alignment Search (MAS).
             2. prior loss: loss between mel-spectrogram and encoder outputs.
             3. flow matching loss: loss between mel-spectrogram and decoder outputs.
 
@@ -177,17 +182,20 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         y_mask = sequence_mask(y_lengths, y_max_length).unsqueeze(1).to(x_mask)
         attn_mask = x_mask.unsqueeze(-1) * y_mask.unsqueeze(2)
 
-        # Use MAS to find most likely alignment `attn` between text and mel-spectrogram
+        if self.use_precomputed_durations:
-        with torch.no_grad():
+            attn = generate_path(durations.squeeze(1), attn_mask.squeeze(1))
-            const = -0.5 * math.log(2 * math.pi) * self.n_feats
+        else:
-            factor = -0.5 * torch.ones(mu_x.shape, dtype=mu_x.dtype, device=mu_x.device)
+            # Use MAS to find most likely alignment `attn` between text and mel-spectrogram
-            y_square = torch.matmul(factor.transpose(1, 2), y**2)
+            with torch.no_grad():
-            y_mu_double = torch.matmul(2.0 * (factor * mu_x).transpose(1, 2), y)
+                const = -0.5 * math.log(2 * math.pi) * self.n_feats
-            mu_square = torch.sum(factor * (mu_x**2), 1).unsqueeze(-1)
+                factor = -0.5 * torch.ones(mu_x.shape, dtype=mu_x.dtype, device=mu_x.device)
-            log_prior = y_square - y_mu_double + mu_square + const
+                y_square = torch.matmul(factor.transpose(1, 2), y**2)
+                y_mu_double = torch.matmul(2.0 * (factor * mu_x).transpose(1, 2), y)
+                mu_square = torch.sum(factor * (mu_x**2), 1).unsqueeze(-1)
+                log_prior = y_square - y_mu_double + mu_square + const
 
-            attn = monotonic_align.maximum_path(log_prior, attn_mask.squeeze(1))
+                attn = monotonic_align.maximum_path(log_prior, attn_mask.squeeze(1))
-            attn = attn.detach()
+                attn = attn.detach()  # b, t_text, T_mel
 
         # Compute loss between predicted log-scaled durations and those obtained from MAS
         # refered to as prior loss in the paper
@@ -228,7 +236,10 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         # Compute loss of the decoder
         diff_loss, _ = self.decoder.compute_loss(x1=y, mask=y_mask, mu=mu_y, spks=spks, cond=cond)
 
-        prior_loss = torch.sum(0.5 * ((y - mu_y) ** 2 + math.log(2 * math.pi)) * y_mask)
+        if self.prior_loss:
-        prior_loss = prior_loss / (torch.sum(y_mask) * self.n_feats)
+            prior_loss = torch.sum(0.5 * ((y - mu_y) ** 2 + math.log(2 * math.pi)) * y_mask)
+            prior_loss = prior_loss / (torch.sum(y_mask) * self.n_feats)
+        else:
+            prior_loss = 0
 
-        return dur_loss, prior_loss, diff_loss
+        return dur_loss, prior_loss, diff_loss, attn

matcha/onnx/export.py

@@ -0,0 +1,181 @@
+import argparse
+import random
+from pathlib import Path
+
+import numpy as np
+import torch
+from lightning import LightningModule
+
+from matcha.cli import VOCODER_URLS, load_matcha, load_vocoder
+
+DEFAULT_OPSET = 15
+
+SEED = 1234
+random.seed(SEED)
+np.random.seed(SEED)
+torch.manual_seed(SEED)
+torch.cuda.manual_seed(SEED)
+torch.backends.cudnn.deterministic = True
+torch.backends.cudnn.benchmark = False
+
+
+class MatchaWithVocoder(LightningModule):
+    def __init__(self, matcha, vocoder):
+        super().__init__()
+        self.matcha = matcha
+        self.vocoder = vocoder
+
+    def forward(self, x, x_lengths, scales, spks=None):
+        mel, mel_lengths = self.matcha(x, x_lengths, scales, spks)
+        wavs = self.vocoder(mel).clamp(-1, 1)
+        lengths = mel_lengths * 256
+        return wavs.squeeze(1), lengths
+
+
+def get_exportable_module(matcha, vocoder, n_timesteps):
+    """
+    Return an appropriate `LighteningModule` and output-node names
+    based on whether the vocoder is embedded in  the final graph
+    """
+
+    def onnx_forward_func(x, x_lengths, scales, spks=None):
+        """
+        Custom forward function for accepting
+        scaler parameters as tensors
+        """
+        # Extract scaler parameters from tensors
+        temperature = scales[0]
+        length_scale = scales[1]
+        output = matcha.synthesise(x, x_lengths, n_timesteps, temperature, spks, length_scale)
+        return output["mel"], output["mel_lengths"]
+
+    # Monkey-patch Matcha's forward function
+    matcha.forward = onnx_forward_func
+
+    if vocoder is None:
+        model, output_names = matcha, ["mel", "mel_lengths"]
+    else:
+        model = MatchaWithVocoder(matcha, vocoder)
+        output_names = ["wav", "wav_lengths"]
+    return model, output_names
+
+
+def get_inputs(is_multi_speaker):
+    """
+    Create dummy inputs for tracing
+    """
+    dummy_input_length = 50
+    x = torch.randint(low=0, high=20, size=(1, dummy_input_length), dtype=torch.long)
+    x_lengths = torch.LongTensor([dummy_input_length])
+
+    # Scales
+    temperature = 0.667
+    length_scale = 1.0
+    scales = torch.Tensor([temperature, length_scale])
+
+    model_inputs = [x, x_lengths, scales]
+    input_names = [
+        "x",
+        "x_lengths",
+        "scales",
+    ]
+
+    if is_multi_speaker:
+        spks = torch.LongTensor([1])
+        model_inputs.append(spks)
+        input_names.append("spks")
+
+    return tuple(model_inputs), input_names
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Export 🍵 Matcha-TTS to ONNX")
+
+    parser.add_argument(
+        "checkpoint_path",
+        type=str,
+        help="Path to the model checkpoint",
+    )
+    parser.add_argument("output", type=str, help="Path to output `.onnx` file")
+    parser.add_argument(
+        "--n-timesteps", type=int, default=5, help="Number of steps to use for reverse diffusion in decoder (default 5)"
+    )
+    parser.add_argument(
+        "--vocoder-name",
+        type=str,
+        choices=list(VOCODER_URLS.keys()),
+        default=None,
+        help="Name of the vocoder to embed in the ONNX graph",
+    )
+    parser.add_argument(
+        "--vocoder-checkpoint-path",
+        type=str,
+        default=None,
+        help="Vocoder checkpoint to embed  in the ONNX graph for an `e2e` like experience",
+    )
+    parser.add_argument("--opset", type=int, default=DEFAULT_OPSET, help="ONNX opset version to use (default 15")
+
+    args = parser.parse_args()
+
+    print(f"[🍵] Loading Matcha checkpoint from {args.checkpoint_path}")
+    print(f"Setting n_timesteps to {args.n_timesteps}")
+
+    checkpoint_path = Path(args.checkpoint_path)
+    matcha = load_matcha(checkpoint_path.stem, checkpoint_path, "cpu")
+
+    if args.vocoder_name or args.vocoder_checkpoint_path:
+        assert (
+            args.vocoder_name and args.vocoder_checkpoint_path
+        ), "Both vocoder_name and vocoder-checkpoint are required when embedding the vocoder in the ONNX graph."
+        vocoder, _ = load_vocoder(args.vocoder_name, args.vocoder_checkpoint_path, "cpu")
+    else:
+        vocoder = None
+
+    is_multi_speaker = matcha.n_spks > 1
+
+    dummy_input, input_names = get_inputs(is_multi_speaker)
+    model, output_names = get_exportable_module(matcha, vocoder, args.n_timesteps)
+
+    # Set dynamic shape for inputs/outputs
+    dynamic_axes = {
+        "x": {0: "batch_size", 1: "time"},
+        "x_lengths": {0: "batch_size"},
+    }
+
+    if vocoder is None:
+        dynamic_axes.update(
+            {
+                "mel": {0: "batch_size", 2: "time"},
+                "mel_lengths": {0: "batch_size"},
+            }
+        )
+    else:
+        print("Embedding the vocoder in the ONNX graph")
+        dynamic_axes.update(
+            {
+                "wav": {0: "batch_size", 1: "time"},
+                "wav_lengths": {0: "batch_size"},
+            }
+        )
+
+    if is_multi_speaker:
+        dynamic_axes["spks"] = {0: "batch_size"}
+
+    # Create the output directory (if not exists)
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+
+    model.to_onnx(
+        args.output,
+        dummy_input,
+        input_names=input_names,
+        output_names=output_names,
+        dynamic_axes=dynamic_axes,
+        opset_version=args.opset,
+        export_params=True,
+        do_constant_folding=True,
+    )
+    print(f"[🍵] ONNX model exported to  {args.output}")
+
+
+if __name__ == "__main__":
+    main()

matcha/onnx/infer.py

@@ -0,0 +1,168 @@
+import argparse
+import os
+import warnings
+from pathlib import Path
+from time import perf_counter
+
+import numpy as np
+import onnxruntime as ort
+import soundfile as sf
+import torch
+
+from matcha.cli import plot_spectrogram_to_numpy, process_text
+
+
+def validate_args(args):
+    assert (
+        args.text or args.file
+    ), "Either text or file must be provided Matcha-T(ea)TTS need sometext to whisk the waveforms."
+    assert args.temperature >= 0, "Sampling temperature cannot be negative"
+    assert args.speaking_rate >= 0, "Speaking rate must be greater than 0"
+    return args
+
+
+def write_wavs(model, inputs, output_dir, external_vocoder=None):
+    if external_vocoder is None:
+        print("The provided model has the vocoder embedded in the graph.\nGenerating waveform directly")
+        t0 = perf_counter()
+        wavs, wav_lengths = model.run(None, inputs)
+        infer_secs = perf_counter() - t0
+        mel_infer_secs = vocoder_infer_secs = None
+    else:
+        print("[🍵] Generating mel using Matcha")
+        mel_t0 = perf_counter()
+        mels, mel_lengths = model.run(None, inputs)
+        mel_infer_secs = perf_counter() - mel_t0
+        print("Generating waveform from mel using external vocoder")
+        vocoder_inputs = {external_vocoder.get_inputs()[0].name: mels}
+        vocoder_t0 = perf_counter()
+        wavs = external_vocoder.run(None, vocoder_inputs)[0]
+        vocoder_infer_secs = perf_counter() - vocoder_t0
+        wavs = wavs.squeeze(1)
+        wav_lengths = mel_lengths * 256
+        infer_secs = mel_infer_secs + vocoder_infer_secs
+
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    for i, (wav, wav_length) in enumerate(zip(wavs, wav_lengths)):
+        output_filename = output_dir.joinpath(f"output_{i + 1}.wav")
+        audio = wav[:wav_length]
+        print(f"Writing audio to {output_filename}")
+        sf.write(output_filename, audio, 22050, "PCM_24")
+
+    wav_secs = wav_lengths.sum() / 22050
+    print(f"Inference seconds: {infer_secs}")
+    print(f"Generated wav seconds: {wav_secs}")
+    rtf = infer_secs / wav_secs
+    if mel_infer_secs is not None:
+        mel_rtf = mel_infer_secs / wav_secs
+        print(f"Matcha RTF: {mel_rtf}")
+    if vocoder_infer_secs is not None:
+        vocoder_rtf = vocoder_infer_secs / wav_secs
+        print(f"Vocoder RTF: {vocoder_rtf}")
+    print(f"Overall RTF: {rtf}")
+
+
+def write_mels(model, inputs, output_dir):
+    t0 = perf_counter()
+    mels, mel_lengths = model.run(None, inputs)
+    infer_secs = perf_counter() - t0
+
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    for i, mel in enumerate(mels):
+        output_stem = output_dir.joinpath(f"output_{i + 1}")
+        plot_spectrogram_to_numpy(mel.squeeze(), output_stem.with_suffix(".png"))
+        np.save(output_stem.with_suffix(".numpy"), mel)
+
+    wav_secs = (mel_lengths * 256).sum() / 22050
+    print(f"Inference seconds: {infer_secs}")
+    print(f"Generated wav seconds: {wav_secs}")
+    rtf = infer_secs / wav_secs
+    print(f"RTF: {rtf}")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=" 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching"
+    )
+    parser.add_argument(
+        "model",
+        type=str,
+        help="ONNX model to use",
+    )
+    parser.add_argument("--vocoder", type=str, default=None, help="Vocoder to use (defaults to None)")
+    parser.add_argument("--text", type=str, default=None, help="Text to synthesize")
+    parser.add_argument("--file", type=str, default=None, help="Text file to synthesize")
+    parser.add_argument("--spk", type=int, default=None, help="Speaker ID")
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=0.667,
+        help="Variance of the x0 noise (default: 0.667)",
+    )
+    parser.add_argument(
+        "--speaking-rate",
+        type=float,
+        default=1.0,
+        help="change the speaking rate, a higher value means slower speaking rate (default: 1.0)",
+    )
+    parser.add_argument("--gpu", action="store_true", help="Use CPU for inference (default: use GPU if available)")
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default=os.getcwd(),
+        help="Output folder to save results (default: current dir)",
+    )
+
+    args = parser.parse_args()
+    args = validate_args(args)
+
+    if args.gpu:
+        providers = ["GPUExecutionProvider"]
+    else:
+        providers = ["CPUExecutionProvider"]
+    model = ort.InferenceSession(args.model, providers=providers)
+
+    model_inputs = model.get_inputs()
+    model_outputs = list(model.get_outputs())
+
+    if args.text:
+        text_lines = args.text.splitlines()
+    else:
+        with open(args.file, encoding="utf-8") as file:
+            text_lines = file.read().splitlines()
+
+    processed_lines = [process_text(0, line, "cpu") for line in text_lines]
+    x = [line["x"].squeeze() for line in processed_lines]
+    # Pad
+    x = torch.nn.utils.rnn.pad_sequence(x, batch_first=True)
+    x = x.detach().cpu().numpy()
+    x_lengths = np.array([line["x_lengths"].item() for line in processed_lines], dtype=np.int64)
+    inputs = {
+        "x": x,
+        "x_lengths": x_lengths,
+        "scales": np.array([args.temperature, args.speaking_rate], dtype=np.float32),
+    }
+    is_multi_speaker = len(model_inputs) == 4
+    if is_multi_speaker:
+        if args.spk is None:
+            args.spk = 0
+            warn = "[!] Speaker ID not provided! Using speaker ID 0"
+            warnings.warn(warn, UserWarning)
+        inputs["spks"] = np.repeat(args.spk, x.shape[0]).astype(np.int64)
+
+    has_vocoder_embedded = model_outputs[0].name == "wav"
+    if has_vocoder_embedded:
+        write_wavs(model, inputs, args.output_dir)
+    elif args.vocoder:
+        external_vocoder = ort.InferenceSession(args.vocoder, providers=providers)
+        write_wavs(model, inputs, args.output_dir, external_vocoder=external_vocoder)
+    else:
+        warn = "[!] A vocoder is not embedded in the graph nor an external vocoder is provided. The mel output will be written as numpy arrays to `*.npy` files in the output directory"
+        warnings.warn(warn, UserWarning)
+        write_mels(model, inputs, args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

matcha/text/__init__.py

@@ -7,6 +7,10 @@ _symbol_to_id = {s: i for i, s in enumerate(symbols)}
 _id_to_symbol = {i: s for i, s in enumerate(symbols)}  # pylint: disable=unnecessary-comprehension
 
 
+class UnknownCleanerException(Exception):
+    pass
+
+
 def text_to_sequence(text, cleaner_names):
     """Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
     Args:
@@ -21,7 +25,7 @@ def text_to_sequence(text, cleaner_names):
     for symbol in clean_text:
         symbol_id = _symbol_to_id[symbol]
         sequence += [symbol_id]
-    return sequence
+    return sequence, clean_text
 
 
 def cleaned_text_to_sequence(cleaned_text):
@@ -48,6 +52,6 @@ def _clean_text(text, cleaner_names):
     for name in cleaner_names:
         cleaner = getattr(cleaners, name)
         if not cleaner:
-            raise Exception("Unknown cleaner: %s" % name)
+            raise UnknownCleanerException(f"Unknown cleaner: {name}")
         text = cleaner(text)
     return text

matcha/text/cleaners.py

@@ -36,9 +36,12 @@ global_phonemizer = phonemizer.backend.EspeakBackend(
 # Regular expression matching whitespace:
 _whitespace_re = re.compile(r"\s+")
 
+# Remove brackets
+_brackets_re = re.compile(r"[\[\]\(\)\{\}]")
+
 # List of (regular expression, replacement) pairs for abbreviations:
 _abbreviations = [
-    (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+    (re.compile(f"\\b{x[0]}\\.", re.IGNORECASE), x[1])
     for x in [
         ("mrs", "misess"),
         ("mr", "mister"),
@@ -72,6 +75,10 @@ def lowercase(text):
     return text.lower()
 
 
+def remove_brackets(text):
+    return re.sub(_brackets_re, "", text)
+
+
 def collapse_whitespace(text):
     return re.sub(_whitespace_re, " ", text)
 
@@ -101,5 +108,37 @@ def english_cleaners2(text):
     text = lowercase(text)
     text = expand_abbreviations(text)
     phonemes = global_phonemizer.phonemize([text], strip=True, njobs=1)[0]
+    # Added in some cases espeak is not removing brackets
+    phonemes = remove_brackets(phonemes)
     phonemes = collapse_whitespace(phonemes)
     return phonemes
+
+
+def ipa_simplifier(text):
+    replacements = [
+        ("ɐ", "ə"),
+        ("ˈə", "ə"),
+        ("ʤ", "dʒ"),
+        ("ʧ", "tʃ"),
+        ("ᵻ", "ɪ"),
+    ]
+    for replacement in replacements:
+        text = text.replace(replacement[0], replacement[1])
+    phonemes = collapse_whitespace(text)
+    return phonemes
+
+
+# I am removing this due to incompatibility with several version of python
+# However, if you want to use it, you can uncomment it
+# and install piper-phonemize with the following command:
+# pip install piper-phonemize
+
+# import piper_phonemize
+# def english_cleaners_piper(text):
+#     """Pipeline for English text, including abbreviation expansion. + punctuation + stress"""
+#     text = convert_to_ascii(text)
+#     text = lowercase(text)
+#     text = expand_abbreviations(text)
+#     phonemes = "".join(piper_phonemize.phonemize_espeak(text=text, voice="en-US")[0])
+#     phonemes = collapse_whitespace(phonemes)
+#     return phonemes

matcha/utils/audio.py

@@ -48,7 +48,7 @@ def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin,
     if torch.max(y) > 1.0:
         print("max value is ", torch.max(y))
 
-    global mel_basis, hann_window  # pylint: disable=global-statement
+    global mel_basis, hann_window  # pylint: disable=global-statement,global-variable-not-assigned
     if f"{str(fmax)}_{str(y.device)}" not in mel_basis:
         mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
         mel_basis[str(fmax) + "_" + str(y.device)] = torch.from_numpy(mel).float().to(y.device)

matcha/utils/data/hificaptain.py

@@ -0,0 +1,148 @@
+#!/usr/bin/env python
+import argparse
+import os
+import sys
+import tempfile
+from pathlib import Path
+
+import torchaudio
+from torch.hub import download_url_to_file
+from tqdm import tqdm
+
+from matcha.utils.data.utils import _extract_zip
+
+URLS = {
+    "en-US": {
+        "female": "https://ast-astrec.nict.go.jp/release/hi-fi-captain/hfc_en-US_F.zip",
+        "male": "https://ast-astrec.nict.go.jp/release/hi-fi-captain/hfc_en-US_M.zip",
+    },
+    "ja-JP": {
+        "female": "https://ast-astrec.nict.go.jp/release/hi-fi-captain/hfc_ja-JP_F.zip",
+        "male": "https://ast-astrec.nict.go.jp/release/hi-fi-captain/hfc_ja-JP_M.zip",
+    },
+}
+
+INFO_PAGE = "https://ast-astrec.nict.go.jp/en/release/hi-fi-captain/"
+
+# On their website they say "We NICT open-sourced Hi-Fi-CAPTAIN",
+# but they use this very-much-not-open-source licence.
+# Dunno if this is open washing or stupidity.
+LICENCE = "CC BY-NC-SA 4.0"
+
+# I'd normally put the citation here. It's on their website.
+# Boo to non-open-source stuff.
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("-s", "--save-dir", type=str, default=None, help="Place to store the downloaded zip files")
+    parser.add_argument(
+        "-r",
+        "--skip-resampling",
+        action="store_true",
+        default=False,
+        help="Skip resampling the data (from 48 to 22.05)",
+    )
+    parser.add_argument(
+        "-l", "--language", type=str, choices=["en-US", "ja-JP"], default="en-US", help="The language to download"
+    )
+    parser.add_argument(
+        "-g",
+        "--gender",
+        type=str,
+        choices=["male", "female"],
+        default="female",
+        help="The gender of the speaker to download",
+    )
+    parser.add_argument(
+        "-o",
+        "--output_dir",
+        type=str,
+        default="data",
+        help="Place to store the converted data. Top-level only, the subdirectory will be created",
+    )
+
+    return parser.parse_args()
+
+
+def process_text(infile, outpath: Path):
+    outmode = "w"
+    if infile.endswith("dev.txt"):
+        outfile = outpath / "valid.txt"
+    elif infile.endswith("eval.txt"):
+        outfile = outpath / "test.txt"
+    else:
+        outfile = outpath / "train.txt"
+        if outfile.exists():
+            outmode = "a"
+    with (
+        open(infile, encoding="utf-8") as inf,
+        open(outfile, outmode, encoding="utf-8") as of,
+    ):
+        for line in inf.readlines():
+            line = line.strip()
+            fileid, rest = line.split(" ", maxsplit=1)
+            outfile = str(outpath / f"{fileid}.wav")
+            of.write(f"{outfile}|{rest}\n")
+
+
+def process_files(zipfile, outpath, resample=True):
+    with tempfile.TemporaryDirectory() as tmpdirname:
+        for filename in tqdm(_extract_zip(zipfile, tmpdirname)):
+            if not filename.startswith(tmpdirname):
+                filename = os.path.join(tmpdirname, filename)
+            if filename.endswith(".txt"):
+                process_text(filename, outpath)
+            elif filename.endswith(".wav"):
+                filepart = filename.rsplit("/", maxsplit=1)[-1]
+                outfile = str(outpath / filepart)
+                arr, sr = torchaudio.load(filename)
+                if resample:
+                    arr = torchaudio.functional.resample(arr, orig_freq=sr, new_freq=22050)
+                torchaudio.save(outfile, arr, 22050)
+            else:
+                continue
+
+
+def main():
+    args = get_args()
+
+    save_dir = None
+    if args.save_dir:
+        save_dir = Path(args.save_dir)
+        if not save_dir.is_dir():
+            save_dir.mkdir()
+
+    if not args.output_dir:
+        print("output directory not specified, exiting")
+        sys.exit(1)
+
+    URL = URLS[args.language][args.gender]
+    dirname = f"hi-fi_{args.language}_{args.gender}"
+
+    outbasepath = Path(args.output_dir)
+    if not outbasepath.is_dir():
+        outbasepath.mkdir()
+    outpath = outbasepath / dirname
+    if not outpath.is_dir():
+        outpath.mkdir()
+
+    resample = True
+    if args.skip_resampling:
+        resample = False
+
+    if save_dir:
+        zipname = URL.rsplit("/", maxsplit=1)[-1]
+        zipfile = save_dir / zipname
+        if not zipfile.exists():
+            download_url_to_file(URL, zipfile, progress=True)
+        process_files(zipfile, outpath, resample)
+    else:
+        with tempfile.NamedTemporaryFile(suffix=".zip", delete=True) as zf:
+            download_url_to_file(URL, zf.name, progress=True)
+            process_files(zf.name, outpath, resample)
+
+
+if __name__ == "__main__":
+    main()

matcha/utils/data/ljspeech.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python
+import argparse
+import random
+import tempfile
+from pathlib import Path
+
+from torch.hub import download_url_to_file
+
+from matcha.utils.data.utils import _extract_tar
+
+URL = "https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2"
+
+INFO_PAGE = "https://keithito.com/LJ-Speech-Dataset/"
+
+LICENCE = "Public domain (LibriVox copyright disclaimer)"
+
+CITATION = """
+@misc{ljspeech17,
+  author       = {Keith Ito and Linda Johnson},
+  title        = {The LJ Speech Dataset},
+  howpublished = {\\url{https://keithito.com/LJ-Speech-Dataset/}},
+  year         = 2017
+}
+"""
+
+
+def decision():
+    return random.random() < 0.98
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("-s", "--save-dir", type=str, default=None, help="Place to store the downloaded zip files")
+    parser.add_argument(
+        "output_dir",
+        type=str,
+        nargs="?",
+        default="data",
+        help="Place to store the converted data (subdirectory LJSpeech-1.1 will be created)",
+    )
+
+    return parser.parse_args()
+
+
+def process_csv(ljpath: Path):
+    if (ljpath / "metadata.csv").exists():
+        basepath = ljpath
+    elif (ljpath / "LJSpeech-1.1" / "metadata.csv").exists():
+        basepath = ljpath / "LJSpeech-1.1"
+    csvpath = basepath / "metadata.csv"
+    wavpath = basepath / "wavs"
+
+    with (
+        open(csvpath, encoding="utf-8") as csvf,
+        open(basepath / "train.txt", "w", encoding="utf-8") as tf,
+        open(basepath / "val.txt", "w", encoding="utf-8") as vf,
+    ):
+        for line in csvf.readlines():
+            line = line.strip()
+            parts = line.split("|")
+            wavfile = str(wavpath / f"{parts[0]}.wav")
+            if decision():
+                tf.write(f"{wavfile}|{parts[1]}\n")
+            else:
+                vf.write(f"{wavfile}|{parts[1]}\n")
+
+
+def main():
+    args = get_args()
+
+    save_dir = None
+    if args.save_dir:
+        save_dir = Path(args.save_dir)
+        if not save_dir.is_dir():
+            save_dir.mkdir()
+
+    outpath = Path(args.output_dir)
+    if not outpath.is_dir():
+        outpath.mkdir()
+
+    if save_dir:
+        tarname = URL.rsplit("/", maxsplit=1)[-1]
+        tarfile = save_dir / tarname
+        if not tarfile.exists():
+            download_url_to_file(URL, str(tarfile), progress=True)
+        _extract_tar(tarfile, outpath)
+        process_csv(outpath)
+    else:
+        with tempfile.NamedTemporaryFile(suffix=".tar.bz2", delete=True) as zf:
+            download_url_to_file(URL, zf.name, progress=True)
+            _extract_tar(zf.name, outpath)
+            process_csv(outpath)
+
+
+if __name__ == "__main__":
+    main()

matcha/utils/data/utils.py

@@ -0,0 +1,53 @@
+# taken from https://github.com/pytorch/audio/blob/main/src/torchaudio/datasets/utils.py
+# Copyright (c) 2017 Facebook Inc. (Soumith Chintala)
+# Licence: BSD 2-Clause
+# pylint: disable=C0123
+
+import logging
+import os
+import tarfile
+import zipfile
+from pathlib import Path
+from typing import Any, List, Optional, Union
+
+_LG = logging.getLogger(__name__)
+
+
+def _extract_tar(from_path: Union[str, Path], to_path: Optional[str] = None, overwrite: bool = False) -> List[str]:
+    if type(from_path) is Path:
+        from_path = str(Path)
+
+    if to_path is None:
+        to_path = os.path.dirname(from_path)
+
+    with tarfile.open(from_path, "r") as tar:
+        files = []
+        for file_ in tar:  # type: Any
+            file_path = os.path.join(to_path, file_.name)
+            if file_.isfile():
+                files.append(file_path)
+                if os.path.exists(file_path):
+                    _LG.info("%s already extracted.", file_path)
+                    if not overwrite:
+                        continue
+            tar.extract(file_, to_path)
+        return files
+
+
+def _extract_zip(from_path: Union[str, Path], to_path: Optional[str] = None, overwrite: bool = False) -> List[str]:
+    if type(from_path) is Path:
+        from_path = str(Path)
+
+    if to_path is None:
+        to_path = os.path.dirname(from_path)
+
+    with zipfile.ZipFile(from_path, "r") as zfile:
+        files = zfile.namelist()
+        for file_ in files:
+            file_path = os.path.join(to_path, file_)
+            if os.path.exists(file_path):
+                _LG.info("%s already extracted.", file_path)
+                if not overwrite:
+                    continue
+            zfile.extract(file_, to_path)
+    return files

matcha/utils/generate_data_statistics.py

@@ -94,6 +94,7 @@ def main():
         cfg["batch_size"] = args.batch_size
         cfg["train_filelist_path"] = str(os.path.join(root_path, cfg["train_filelist_path"]))
         cfg["valid_filelist_path"] = str(os.path.join(root_path, cfg["valid_filelist_path"]))
+        cfg["load_durations"] = False
 
     text_mel_datamodule = TextMelDataModule(**cfg)
     text_mel_datamodule.setup()
@@ -101,10 +102,8 @@ def main():
     log.info("Dataloader loaded! Now computing stats...")
     params = compute_data_statistics(data_loader, cfg["n_feats"])
     print(params)
-    json.dump(
+    with open(output_file, "w", encoding="utf-8") as dumpfile:
-        params,
+        json.dump(params, dumpfile)
-        open(output_file, "w"),
-    )
 
 
 if __name__ == "__main__":

matcha/utils/get_durations_from_trained_model.py

@@ -0,0 +1,195 @@
+r"""
+The file creates a pickle file where the values needed for loading of dataset is stored and the model can load it
+when needed.
+
+Parameters from hparam.py will be used
+"""
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import lightning
+import numpy as np
+import rootutils
+import torch
+from hydra import compose, initialize
+from omegaconf import open_dict
+from torch import nn
+from tqdm.auto import tqdm
+
+from matcha.cli import get_device
+from matcha.data.text_mel_datamodule import TextMelDataModule
+from matcha.models.matcha_tts import MatchaTTS
+from matcha.utils.logging_utils import pylogger
+from matcha.utils.utils import get_phoneme_durations
+
+log = pylogger.get_pylogger(__name__)
+
+
+def save_durations_to_folder(
+    attn: torch.Tensor, x_length: int, y_length: int, filepath: str, output_folder: Path, text: str
+):
+    durations = attn.squeeze().sum(1)[:x_length].numpy()
+    durations_json = get_phoneme_durations(durations, text)
+    output = output_folder / Path(filepath).name.replace(".wav", ".npy")
+    with open(output.with_suffix(".json"), "w", encoding="utf-8") as f:
+        json.dump(durations_json, f, indent=4, ensure_ascii=False)
+
+    np.save(output, durations)
+
+
+@torch.inference_mode()
+def compute_durations(data_loader: torch.utils.data.DataLoader, model: nn.Module, device: torch.device, output_folder):
+    """Generate durations from the model for each datapoint and save it in a folder
+
+    Args:
+        data_loader (torch.utils.data.DataLoader): Dataloader
+        model (nn.Module): MatchaTTS model
+        device (torch.device): GPU or CPU
+    """
+
+    for batch in tqdm(data_loader, desc="🍵 Computing durations 🍵:"):
+        x, x_lengths = batch["x"], batch["x_lengths"]
+        y, y_lengths = batch["y"], batch["y_lengths"]
+        spks = batch["spks"]
+        x = x.to(device)
+        y = y.to(device)
+        x_lengths = x_lengths.to(device)
+        y_lengths = y_lengths.to(device)
+        spks = spks.to(device) if spks is not None else None
+
+        _, _, _, attn = model(
+            x=x,
+            x_lengths=x_lengths,
+            y=y,
+            y_lengths=y_lengths,
+            spks=spks,
+        )
+        attn = attn.cpu()
+        for i in range(attn.shape[0]):
+            save_durations_to_folder(
+                attn[i],
+                x_lengths[i].item(),
+                y_lengths[i].item(),
+                batch["filepaths"][i],
+                output_folder,
+                batch["x_texts"][i],
+            )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "-i",
+        "--input-config",
+        type=str,
+        default="ljspeech.yaml",
+        help="The name of the yaml config file under configs/data",
+    )
+
+    parser.add_argument(
+        "-b",
+        "--batch-size",
+        type=int,
+        default="32",
+        help="Can have increased batch size for faster computation",
+    )
+
+    parser.add_argument(
+        "-f",
+        "--force",
+        action="store_true",
+        default=False,
+        required=False,
+        help="force overwrite the file",
+    )
+    parser.add_argument(
+        "-c",
+        "--checkpoint_path",
+        type=str,
+        required=True,
+        help="Path to the checkpoint file to load the model from",
+    )
+
+    parser.add_argument(
+        "-o",
+        "--output-folder",
+        type=str,
+        default=None,
+        help="Output folder to save the data statistics",
+    )
+
+    parser.add_argument(
+        "--cpu", action="store_true", help="Use CPU for inference, not recommended (default: use GPU if available)"
+    )
+
+    args = parser.parse_args()
+
+    with initialize(version_base="1.3", config_path="../../configs/data"):
+        cfg = compose(config_name=args.input_config, return_hydra_config=True, overrides=[])
+
+    root_path = rootutils.find_root(search_from=__file__, indicator=".project-root")
+
+    with open_dict(cfg):
+        del cfg["hydra"]
+        del cfg["_target_"]
+        cfg["seed"] = 1234
+        cfg["batch_size"] = args.batch_size
+        cfg["train_filelist_path"] = str(os.path.join(root_path, cfg["train_filelist_path"]))
+        cfg["valid_filelist_path"] = str(os.path.join(root_path, cfg["valid_filelist_path"]))
+        cfg["load_durations"] = False
+
+    if args.output_folder is not None:
+        output_folder = Path(args.output_folder)
+    else:
+        output_folder = Path(cfg["train_filelist_path"]).parent / "durations"
+
+    print(f"Output folder set to: {output_folder}")
+
+    if os.path.exists(output_folder) and not args.force:
+        print("Folder already exists. Use -f to force overwrite")
+        sys.exit(1)
+
+    output_folder.mkdir(parents=True, exist_ok=True)
+
+    print(f"Preprocessing: {cfg['name']} from training filelist: {cfg['train_filelist_path']}")
+    print("Loading model...")
+    device = get_device(args)
+    model = MatchaTTS.load_from_checkpoint(args.checkpoint_path, map_location=device)
+
+    text_mel_datamodule = TextMelDataModule(**cfg)
+    text_mel_datamodule.setup()
+    try:
+        print("Computing stats for training set if exists...")
+        train_dataloader = text_mel_datamodule.train_dataloader()
+        compute_durations(train_dataloader, model, device, output_folder)
+    except lightning.fabric.utilities.exceptions.MisconfigurationException:
+        print("No training set found")
+
+    try:
+        print("Computing stats for validation set if exists...")
+        val_dataloader = text_mel_datamodule.val_dataloader()
+        compute_durations(val_dataloader, model, device, output_folder)
+    except lightning.fabric.utilities.exceptions.MisconfigurationException:
+        print("No validation set found")
+
+    try:
+        print("Computing stats for test set if exists...")
+        test_dataloader = text_mel_datamodule.test_dataloader()
+        compute_durations(test_dataloader, model, device, output_folder)
+    except lightning.fabric.utilities.exceptions.MisconfigurationException:
+        print("No test set found")
+
+    print(f"[+] Done! Data statistics saved to: {output_folder}")
+
+
+if __name__ == "__main__":
+    # Helps with generating durations for the dataset to train other architectures
+    # that cannot learn to align due to limited size of dataset
+    # Example usage:
+    # python python matcha/utils/get_durations_from_trained_model.py -i ljspeech.yaml -c pretrained_model
+    # This will create a folder in data/processed_data/durations/ljspeech with the durations
+    main()

matcha/utils/model.py

@@ -7,15 +7,17 @@ import torch
 def sequence_mask(length, max_length=None):
     if max_length is None:
         max_length = length.max()
-    x = torch.arange(int(max_length), dtype=length.dtype, device=length.device)
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
     return x.unsqueeze(0) < length.unsqueeze(1)
 
 
 def fix_len_compatibility(length, num_downsamplings_in_unet=2):
-    while True:
+    factor = torch.scalar_tensor(2).pow(num_downsamplings_in_unet)
-        if length % (2**num_downsamplings_in_unet) == 0:
+    length = (length / factor).ceil() * factor
-            return length
+    if not torch.onnx.is_in_onnx_export():
-        length += 1
+        return length.int().item()
+    else:
+        return length
 
 
 def convert_pad_shape(pad_shape):

matcha/utils/rich_utils.py

@@ -72,7 +72,7 @@ def print_config_tree(
 
     # save config tree to file
     if save_to_file:
-        with open(Path(cfg.paths.output_dir, "config_tree.log"), "w") as file:
+        with open(Path(cfg.paths.output_dir, "config_tree.log"), "w", encoding="utf-8") as file:
             rich.print(tree, file=file)
 
 
@@ -97,5 +97,5 @@ def enforce_tags(cfg: DictConfig, save_to_file: bool = False) -> None:
         log.info(f"Tags: {cfg.tags}")
 
     if save_to_file:
-        with open(Path(cfg.paths.output_dir, "tags.log"), "w") as file:
+        with open(Path(cfg.paths.output_dir, "tags.log"), "w", encoding="utf-8") as file:
             rich.print(cfg.tags, file=file)

matcha/utils/utils.py

@@ -2,6 +2,7 @@ import os
 import sys
 import warnings
 from importlib.util import find_spec
+from math import ceil
 from pathlib import Path
 from typing import Any, Callable, Dict, Tuple
 
@@ -115,7 +116,7 @@ def get_metric_value(metric_dict: Dict[str, Any], metric_name: str) -> float:
         return None
 
     if metric_name not in metric_dict:
-        raise Exception(
+        raise ValueError(
             f"Metric value not found! <metric_name={metric_name}>\n"
             "Make sure metric name logged in LightningModule is correct!\n"
             "Make sure `optimized_metric` name in `hparams_search` config is correct!"
@@ -205,13 +206,54 @@ def get_user_data_dir(appname="matcha_tts"):
     return final_path
 
 
-def assert_model_downloaded(checkpoint_path, url, use_wget=False):
+def assert_model_downloaded(checkpoint_path, url, use_wget=True):
     if Path(checkpoint_path).exists():
         log.debug(f"[+] Model already present at {checkpoint_path}!")
+        print(f"[+] Model already present at {checkpoint_path}!")
         return
     log.info(f"[-] Model not found at {checkpoint_path}! Will download it")
+    print(f"[-] Model not found at {checkpoint_path}! Will download it")
     checkpoint_path = str(checkpoint_path)
     if not use_wget:
         gdown.download(url=url, output=checkpoint_path, quiet=False, fuzzy=True)
     else:
         wget.download(url=url, out=checkpoint_path)
+
+
+def get_phoneme_durations(durations, phones):
+    prev = durations[0]
+    merged_durations = []
+    # Convolve with stride 2
+    for i in range(1, len(durations), 2):
+        if i == len(durations) - 2:
+            # if it is last take full value
+            next_half = durations[i + 1]
+        else:
+            next_half = ceil(durations[i + 1] / 2)
+
+        curr = prev + durations[i] + next_half
+        prev = durations[i + 1] - next_half
+        merged_durations.append(curr)
+
+    assert len(phones) == len(merged_durations)
+    assert len(merged_durations) == (len(durations) - 1) // 2
+
+    merged_durations = torch.cumsum(torch.tensor(merged_durations), 0, dtype=torch.long)
+    start = torch.tensor(0)
+    duration_json = []
+    for i, duration in enumerate(merged_durations):
+        duration_json.append(
+            {
+                phones[i]: {
+                    "starttime": start.item(),
+                    "endtime": duration.item(),
+                    "duration": duration.item() - start.item(),
+                }
+            }
+        )
+        start = duration
+
+    assert list(duration_json[-1].values())[0]["endtime"] == sum(
+        durations
+    ), f"{list(duration_json[-1].values())[0]['endtime'],  sum(durations)}"
+    return duration_json

requirements.txt

@@ -35,10 +35,10 @@ torchaudio
 matplotlib
 pandas
 conformer==0.3.2
-diffusers==0.21.1
+diffusers # developed using version ==0.25.0
 notebook
 ipywidgets
-gradio
+gradio==3.43.2
 gdown
 wget
 seaborn

setup.py

@@ -16,9 +16,16 @@ with open("README.md", encoding="utf-8") as readme_file:
     README = readme_file.read()
 
 cwd = os.path.dirname(os.path.abspath(__file__))
-with open(os.path.join(cwd, "matcha", "VERSION")) as fin:
+with open(os.path.join(cwd, "matcha", "VERSION"), encoding="utf-8") as fin:
     version = fin.read().strip()
 
+
+def get_requires():
+    requirements = os.path.join(os.path.dirname(__file__), "requirements.txt")
+    with open(requirements, encoding="utf-8") as reqfile:
+        return [str(r).strip() for r in reqfile]
+
+
 setup(
     name="matcha-tts",
     version=version,
@@ -28,7 +35,7 @@ setup(
     author="Shivam Mehta",
     author_email="shivam.mehta25@gmail.com",
     url="https://shivammehta25.github.io/Matcha-TTS",
-    install_requires=[str(r) for r in open(os.path.join(os.path.dirname(__file__), "requirements.txt"))],
+    install_requires=get_requires(),
     include_dirs=[numpy.get_include()],
     include_package_data=True,
     packages=find_packages(exclude=["tests", "tests/*", "examples", "examples/*"]),
@@ -38,6 +45,7 @@ setup(
             "matcha-data-stats=matcha.utils.generate_data_statistics:main",
             "matcha-tts=matcha.cli:cli",
             "matcha-tts-app=matcha.app:main",
+            "matcha-tts-get-durations=matcha.utils.get_durations_from_trained_model:main",
         ]
     },
     ext_modules=cythonize(exts, language_level=3),

synthesis.ipynb

@@ -19,7 +19,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "id": "148f4bc0-c28e-4670-9a5e-4c7928ab8992",
    "metadata": {},
    "outputs": [
@@ -192,7 +192,7 @@
    "source": [
     "@torch.inference_mode()\n",
     "def process_text(text: str):\n",
-    "    x = torch.tensor(intersperse(text_to_sequence(text, ['english_cleaners2']), 0),dtype=torch.long, device=device)[None]\n",
+    "    x = torch.tensor(intersperse(text_to_sequence(text, ['english_cleaners2'])[0], 0),dtype=torch.long, device=device)[None]\n",
     "    x_lengths = torch.tensor([x.shape[-1]],dtype=torch.long, device=device)\n",
     "    x_phones = sequence_to_text(x.squeeze(0).tolist())\n",
     "    return {\n",