Update README.md

ljspeech/hificaptain from #99

.gitignore

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

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>

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
+train_filelist_path: data/hi-fi_en-US_female/train.txt
-valid_filelist_path: data/filelists/hi-fi-captain-en-us-female_val.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

data

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

matcha/VERSION

@@ -1 +1 @@
-0.0.6.0
+0.0.7.2

matcha/cli.py

@@ -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()
 
@@ -336,7 +336,7 @@ def batched_synthesis(args, device, model, vocoder, denoiser, texts, 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}")
@@ -377,7 +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

@@ -234,9 +234,9 @@ 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)

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/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
@@ -106,6 +106,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
                 # Lengths of mel spectrograms
                 "rtf": float,
                 # Real-time factor
+            }
         """
         # For RTF computation
         t = dt.datetime.now()
@@ -152,7 +153,7 @@ class MatchaTTS(BaseLightningClass):  # 🍵
     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.
 

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:
@@ -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)
 
@@ -101,10 +108,26 @@ 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:

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

@@ -102,10 +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/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)

requirements.txt

@@ -35,7 +35,7 @@ torchaudio
 matplotlib
 pandas
 conformer==0.3.2
-diffusers==0.25.0
+diffusers # developed using version ==0.25.0
 notebook
 ipywidgets
 gradio==3.43.2

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/*"]),