Update README.md

ljspeech/hificaptain from #99

.gitignore

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

.pre-commit-config.yaml

@@ -1,5 +1,5 @@
 default_language_version:
-  python: python3.10
+  python: python3.11
 
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks

README.md

@@ -10,7 +10,7 @@
 [![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>
@@ -252,6 +252,43 @@ python3 -m matcha.onnx.infer model.onnx --text "hey" --output-dir ./outputs --vo
 
 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:

configs/data/hi-fi_en-US_female.yaml

@@ -5,8 +5,8 @@ defaults:
 # 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
+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

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
-valid_filelist_path: data/filelists/ljs_audio_text_val_filelist.txt
+train_filelist_path: data/LJSpeech-1.1/train.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/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

@@ -13,3 +13,4 @@ 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 @@
-0.0.5.1
+0.0.7.2

matcha/cli.py

@@ -48,7 +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]
@@ -114,10 +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()
 
@@ -326,16 +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}")
@@ -376,7 +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])

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,

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/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
@@ -35,6 +35,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         optimizer=None,
         scheduler=None,
         prior_loss=True,
+        use_precomputed_durations=False,
     ):
         super().__init__()
 
@@ -46,6 +47,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         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)
@@ -104,6 +106,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
                 # Lengths of mel spectrograms
                 "rtf": float,
                 # Real-time factor
+            }
         """
         # For RTF computation
         t = dt.datetime.now()
@@ -147,10 +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.
 
@@ -179,17 +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
-        with torch.no_grad():
-            const = -0.5 * math.log(2 * math.pi) * self.n_feats
-            factor = -0.5 * torch.ones(mu_x.shape, dtype=mu_x.dtype, device=mu_x.device)
-            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
+        if self.use_precomputed_durations:
+            attn = generate_path(durations.squeeze(1), attn_mask.squeeze(1))
+        else:
+            # Use MAS to find most likely alignment `attn` between text and mel-spectrogram
+            with torch.no_grad():
+                const = -0.5 * math.log(2 * math.pi) * self.n_feats
+                factor = -0.5 * torch.ones(mu_x.shape, dtype=mu_x.dtype, device=mu_x.device)
+                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 = attn.detach()
+                attn = monotonic_align.maximum_path(log_prior, attn_mask.squeeze(1))
+                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
@@ -236,4 +242,4 @@ class MatchaTTS(BaseLightningClass):  # 🍵
         else:
             prior_loss = 0
 
-        return dur_loss, prior_loss, diff_loss
+        return dur_loss, prior_loss, diff_loss, attn

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

@@ -15,7 +15,6 @@ 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
@@ -37,9 +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"),
@@ -73,6 +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)
 
@@ -102,15 +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 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)
+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(
-        params,
-        open(output_file, "w"),
-    )
+    with open(output_file, "w", encoding="utf-8") as dumpfile:
+        json.dump(params, dumpfile)
 
 
 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/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
 
@@ -217,3 +218,42 @@ def assert_model_downloaded(checkpoint_path, url, use_wget=True):
         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

@@ -35,11 +35,10 @@ torchaudio
 matplotlib
 pandas
 conformer==0.3.2
-diffusers==0.25.0
+diffusers # developed using version ==0.25.0
 notebook
 ipywidgets
-gradio
+gradio==3.43.2
 gdown
 wget
 seaborn
-piper_phonemize

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",