Adding ICASSP 2024

Removing gdown for HifiGAN checkpoints too

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
 
 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,
-                       Exception
+overgeneral-exceptions=builtins.BaseException,
+                       builtins.Exception

README.md

@@ -17,7 +17,7 @@
 
 </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](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:
 
@@ -26,13 +26,13 @@ We propose 🍵 Matcha-TTS, a new approach to non-autoregressive neural TTS, tha
 - Sounds highly natural
 - Is very fast to synthesise from
 
-Check out our [demo page](https://shivammehta25.github.io/Matcha-TTS) and read [our arXiv preprint](https://arxiv.org/abs/2309.03199) for more details.
+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.
 
 [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).
 
-## Watch the teaser
+## Teaser video
 
 [![Watch the video](https://img.youtube.com/vi/xmvJkz3bqw0/hqdefault.jpg)](https://youtu.be/xmvJkz3bqw0)
 
@@ -257,11 +257,11 @@ This will write `.wav` audio files to the output directory.
 If you use our code or otherwise find this work useful, please cite our paper:
 
 ```text
-@article{mehta2023matcha,
-  title={Matcha-TTS: A fast TTS architecture with conditional flow matching},
+@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},
-  journal={arXiv preprint arXiv:2309.03199},
-  year={2023}
+  booktitle={Proc. ICASSP},
+  year={2024}
 }
 ```
 
@@ -269,7 +269,7 @@ If you use our code or otherwise find this work useful, please cite our paper:
 
 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 code 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
 - [Hugging Face Diffusers](https://huggingface.co/): For their awesome diffusers library and its components

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/filelists/hi-fi-captain-en-us-female_train.txt
+valid_filelist_path: data/filelists/hi-fi-captain-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/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/model/matcha.yaml

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

matcha/VERSION

@@ -1 +1 @@
-0.0.4
+0.0.5.1

matcha/app.py

@@ -29,8 +29,15 @@ args = Namespace(
 
 CURRENTLY_LOADED_MODEL = args.model
 
-MATCHA_TTS_LOC = lambda x: LOCATION / f"{x}.ckpt"  # noqa: E731
-VOCODER_LOC = lambda x: LOCATION / f"{x}"  # noqa: E731
+
+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"
 RADIO_OPTIONS = {
     "Multi Speaker (VCTK)": {

matcha/cli.py

@@ -18,13 +18,13 @@ 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_vctk": "https://drive.google.com/file/d/1enuxmfslZciWGAl63WGh2ekVo00FYuQ9/view?usp=drive_link",
+    "matcha_ljspeech": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_ljspeech.ckpt",
+    "matcha_vctk": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_vctk.ckpt",
 }
 
 VOCODER_URLS = {
-    "hifigan_T2_v1": "https://drive.google.com/file/d/14NENd4equCBLyyCSke114Mv6YR_j_uFs/view?usp=drive_link",
-    "hifigan_univ_v1": "https://drive.google.com/file/d/1qpgI41wNXFcH-iKq1Y42JlBC9j0je8PW/view?usp=drive_link",
+    "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
 }
 
 MULTISPEAKER_MODEL = {
@@ -63,7 +63,7 @@ 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
 
@@ -140,7 +140,7 @@ def validate_args(args):
 
     if args.checkpoint_path is None:
         # When using pretrained models
-        if args.model in SINGLESPEAKER_MODEL.keys():
+        if args.model in SINGLESPEAKER_MODEL:
             args = validate_args_for_single_speaker_model(args)
 
         if args.model in MULTISPEAKER_MODEL:

matcha/models/baselightningmodule.py

@@ -81,7 +81,7 @@ 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/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
-        while steps <= len(t_span) - 1:
+        for step in range(1, len(t_span)):
             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:
-                dt = t_span[steps + 1] - t
-            steps += 1
+            if step < len(t_span) - 1:
+                dt = t_span[step + 1] - t
 
         return sol[-1]
 

matcha/models/matcha_tts.py

@@ -34,6 +34,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         out_size,
         optimizer=None,
         scheduler=None,
+        prior_loss=True,
     ):
         super().__init__()
 
@@ -44,6 +45,7 @@ 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
 
         if n_spks > 1:
             self.spk_emb = torch.nn.Embedding(n_spks, spk_emb_dim)
@@ -228,7 +230,10 @@ 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)
 
+        if self.prior_loss:
             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

matcha/text/cleaners.py

@@ -15,6 +15,7 @@ import logging
 import re
 
 import phonemizer
+import piper_phonemize
 from unidecode import unidecode
 
 # To avoid excessive logging we set the log level of the phonemizer package to Critical
@@ -103,3 +104,13 @@ def english_cleaners2(text):
     phonemes = global_phonemizer.phonemize([text], strip=True, njobs=1)[0]
     phonemes = collapse_whitespace(phonemes)
     return phonemes
+
+
+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/utils.py

@@ -115,7 +115,7 @@ 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!"
@@ -205,11 +205,13 @@ 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)

requirements.txt

@@ -35,10 +35,11 @@ torchaudio
 matplotlib
 pandas
 conformer==0.3.2
-diffusers==0.21.3
+diffusers==0.25.0
 notebook
 ipywidgets
 gradio
 gdown
 wget
 seaborn
+piper_phonemize