libritts-aligned.py

"""LibriTTS dataset with forced alignments."""

import os
from pathlib import Path
import hashlib
import pickle

import datasets
import pandas as pd
import numpy as np
from alignments.datasets.libritts import LibrittsDataset
from tqdm.contrib.concurrent import process_map
from tqdm.auto import tqdm
from multiprocessing import cpu_count
from phones.convert import Converter
import torchaudio
import torchaudio.transforms as AT

logger = datasets.logging.get_logger(__name__)

_PHONESET = "arpabet"

_VERBOSE = os.environ.get("LIBRITTS_VERBOSE", True)
_MAX_WORKERS = os.environ.get("LIBRITTS_MAX_WORKERS", cpu_count())
_PATH = os.environ.get("LIBRITTS_PATH", os.environ.get("HF_DATASETS_CACHE", None))
if _PATH is not None and not os.path.exists(_PATH):
    os.makedirs(_PATH)

_VERSION = "1.0.1"

_CITATION = """\
@article{zen2019libritts,
  title={LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech},
  author={Zen, Heiga and Dang, Viet and Clark, Rob and Zhang, Yu and Weiss, Ron J and Jia, Ye and Chen, Zhifeng and Wu, Yonghui},
  journal={Interspeech},
  year={2019}
}
@article{https://doi.org/10.48550/arxiv.2211.16049,
  author = {Minixhofer, Christoph and Klejch, Ondřej and Bell, Peter},
  title = {Evaluating and reducing the distance between synthetic and real speech distributions},
  year = {2022}
}
"""

_DESCRIPTION = """\
Dataset used for loading TTS spectrograms and waveform audio with alignments and a number of configurable "measures", which are extracted from the raw audio.
"""

_URL = "https://www.openslr.org/resources/60/"
_URLS = {
    "dev-clean": _URL + "dev-clean.tar.gz",
    "dev-other": _URL + "dev-other.tar.gz",
    "test-clean": _URL + "test-clean.tar.gz",
    "test-other": _URL + "test-other.tar.gz",
    "train-clean-100": _URL + "train-clean-100.tar.gz",
    "train-clean-360": _URL + "train-clean-360.tar.gz",
    "train-other-500": _URL + "train-other-500.tar.gz",
}


class LibriTTSAlignConfig(datasets.BuilderConfig):
    """BuilderConfig for LibriTTSAlign."""

    def __init__(self, sampling_rate=22050, hop_length=256, win_length=1024, **kwargs):
        """BuilderConfig for LibriTTSAlign.

        Args:
          **kwargs: keyword arguments forwarded to super.
        """
        super(LibriTTSAlignConfig, self).__init__(**kwargs)

        self.sampling_rate = sampling_rate
        self.hop_length = hop_length
        self.win_length = win_length
        
        if _PATH is None:
            raise ValueError("Please set the environment variable LIBRITTS_PATH to point to the LibriTTS dataset directory.")
        elif _PATH == os.environ.get("HF_DATASETS_CACHE", None):
            logger.warning("Please set the environment variable LIBRITTS_PATH to point to the LibriTTS dataset directory. Using HF_DATASETS_CACHE as a fallback.")

class LibriTTSAlign(datasets.GeneratorBasedBuilder):
    """LibriTTSAlign dataset."""

    BUILDER_CONFIGS = [
        LibriTTSAlignConfig(
            name="libritts",
            version=datasets.Version(_VERSION, ""),
        ),
    ]

    def _info(self):
        features = {
            "id": datasets.Value("string"),
            "speaker": datasets.Value("string"),
            "text": datasets.Value("string"),
            "start": datasets.Value("float32"),
            "end": datasets.Value("float32"),
            # phone features
            "phones": datasets.Sequence(datasets.Value("string")),
            "phone_durations": datasets.Sequence(datasets.Value("int32")),
            # audio feature
            "audio": datasets.Value("string")
        }

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(features),
            supervised_keys=None,
            homepage="https://github.com/MiniXC/MeasureCollator",
            citation=_CITATION,
            task_templates=None,
        )

    def _split_generators(self, dl_manager):
        ds_dict = {}
        for name, url in _URLS.items():
            ds_dict[name] = self._create_alignments_ds(name, url)
        splits = [
            datasets.SplitGenerator(
                name=key.replace("-", "."),
                gen_kwargs={"ds": self._create_data(value)}
            ) 
            for key, value in ds_dict.items()
        ]
        # dataframe with all data
        data_train = self._create_data([ds_dict["train-clean-100"], ds_dict["train-clean-360"], ds_dict["train-other-500"]])
        data_dev = self._create_data([ds_dict["dev-clean"], ds_dict["dev-other"]])
        data_test = self._create_data([ds_dict["test-clean"], ds_dict["test-other"]])
        data_all = pd.concat([data_train, data_dev, data_test])
        splits += [
            datasets.SplitGenerator(
                name="train.all",
                gen_kwargs={
                    "ds": data_all,
                }
            ),
            datasets.SplitGenerator(
                name="dev.all",
                gen_kwargs={
                    "ds": data_dev,
                }
            ),
            datasets.SplitGenerator(
                name="test.all",
                gen_kwargs={
                    "ds": data_test,
                }
            ),
        ]
        # move last row for each speaker from data_all to dev dataframe
        data_dev = data_all.copy()
        data_dev = data_dev.sort_values(by=["speaker", "audio"])
        data_dev = data_dev.groupby("speaker").tail(1)
        data_dev = data_dev.reset_index()
        # remove last row for each speaker from data_all
        data_all = data_all[~data_all["audio"].isin(data_dev["audio"])]
        splits += [
            datasets.SplitGenerator(
                name="train",
                gen_kwargs={
                    "ds": data_all,
                }
            ),
            datasets.SplitGenerator(
                name="dev",
                gen_kwargs={
                    "ds": data_dev,
                }
            ),
        ]
        self.alignments_ds = None
        self.data = None
        return splits

    def _create_alignments_ds(self, name, url):
        self.empty_textgrids = 0
        ds_hash = hashlib.md5(os.path.join(_PATH, f"{name}-alignments").encode()).hexdigest()
        pkl_path = os.path.join(_PATH, f"{ds_hash}.pkl")
        if os.path.exists(pkl_path):
            ds = pickle.load(open(pkl_path, "rb"))
        else:
            tgt_dir = os.path.join(_PATH, f"{name}-alignments")
            src_dir = os.path.join(_PATH, f"{name}-data")
            if os.path.exists(tgt_dir):
                src_dir = None
                url = None
            if os.path.exists(src_dir):
                url = None
            ds = LibrittsDataset(
                target_directory=tgt_dir,
                source_directory=src_dir,
                source_url=url,
                verbose=_VERBOSE,
                tmp_directory=os.path.join(_PATH, f"{name}-tmp"),
                chunk_size=1000,
            )
            pickle.dump(ds, open(pkl_path, "wb"))
        return ds, ds_hash

    def _create_data(self, data):
        entries = []
        self.phone_cache = {}
        self.phone_converter = Converter()
        if not isinstance(data, list):
            data = [data]
        hashes = [ds_hash for ds, ds_hash in data]
        ds = [ds for ds, ds_hash in data]
        self.ds = ds
        del data
        for i, ds in enumerate(ds):
            if os.path.exists(os.path.join(_PATH, f"{hashes[i]}-entries.pkl")):
                add_entries = pickle.load(open(os.path.join(_PATH, f"{hashes[i]}-entries.pkl"), "rb"))
            else:
                add_entries = [
                    entry
                    for entry in process_map(
                        self._create_entry,
                        zip([i] * len(ds), np.arange(len(ds))),
                        chunksize=10_000,
                        max_workers=_MAX_WORKERS,
                        desc=f"processing dataset {hashes[i]}",
                        tqdm_class=tqdm,
                    )
                    if entry is not None
                ]
                pickle.dump(add_entries, open(os.path.join(_PATH, f"{hashes[i]}-entries.pkl"), "wb"))
            entries += add_entries
        if self.empty_textgrids > 0:
            logger.warning(f"Found {self.empty_textgrids} empty textgrids")
        return pd.DataFrame(
            entries,
            columns=[
                "phones",
                "duration",
                "start",
                "end",
                "audio",
                "speaker",
                "text",
                "basename",
            ],
        )
        del self.ds, self.phone_cache, self.phone_converter

    def _create_entry(self, dsi_idx):
        dsi, idx = dsi_idx
        item = self.ds[dsi][idx]
        start, end = item["phones"][0][0], item["phones"][-1][1]

        phones = []
        durations = []

        for i, p in enumerate(item["phones"]):
            s, e, phone = p
            phone.replace("ˌ", "")
            r_phone = phone.replace("0", "").replace("1", "")
            if len(r_phone) > 0:
                phone = r_phone
            if "[" not in phone:
                o_phone = phone
                if o_phone not in self.phone_cache:
                    phone = self.phone_converter(
                        phone, _PHONESET, lang=None
                    )[0]
                    self.phone_cache[o_phone] = phone
                phone = self.phone_cache[o_phone]
            phones.append(phone)
            durations.append(
                int(
                    np.round(e * self.config.sampling_rate / self.config.hop_length)
                    - np.round(s * self.config.sampling_rate / self.config.hop_length)
                )
            )

        if start >= end:
            self.empty_textgrids += 1
            return None

        return (
            phones,
            durations,
            start,
            end,
            item["wav"],
            str(item["speaker"]).split("/")[-1],
            item["transcript"],
            Path(item["wav"]).name,
        )

    def _generate_examples(self, ds):
        j = 0
        for i, row in ds.iterrows():
            # 10kB is the minimum size of a wav file for our purposes
            if Path(row["audio"]).stat().st_size >= 10_000:
                if len(row["phones"]) < 384:
                    result = {
                        "id": row["basename"],
                        "speaker": row["speaker"],
                        "text": row["text"],
                        "start": row["start"],
                        "end": row["end"],
                        "phones": row["phones"],
                        "phone_durations": row["duration"],
                        "audio": str(row["audio"]),
                    }
                    yield j, result
                    j += 1