loader.py

"""HuggingFace loading script for the JamALT dataset."""


import csv
from dataclasses import dataclass
import json
import os
from pathlib import Path
from typing import Optional

import datasets


_VERSION = "1.1.0"


_CITATION = """\
@misc{cifka-2023-jam-alt,
  author       = {Ond\v{r}ej C\'ifka and
                  Constantinos Dimitriou and
                  {Cheng-i} Wang and
                  Hendrik Schreiber and
                  Luke Miner and
                  Fabian-Robert St\"oter},
  title        = {{Jam-ALT}: A Formatting-Aware Lyrics Transcription Benchmark},
  eprint       = {arXiv:2311.13987},
  year         = 2023
}
@inproceedings{durand-2023-contrastive,
  author={Durand, Simon and Stoller, Daniel and Ewert, Sebastian},
  booktitle={2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Contrastive Learning-Based Audio to Lyrics Alignment for Multiple Languages}, 
  year={2023},
  pages={1-5},
  address={Rhodes Island, Greece},
  doi={10.1109/ICASSP49357.2023.10096725}
}
"""

_DESCRIPTION = """\
Jam-ALT: A formatting-aware lyrics transcription benchmark.
"""

_HOMEPAGE = "https://audioshake.github.io/jam-alt"

_METADATA_FILENAME = "metadata.csv"


_LANGUAGE_NAME_TO_CODE = {
    "English": "en",
    "French": "fr",
    "German": "de",
    "Spanish": "es",
}


@dataclass
class JamAltBuilderConfig(datasets.BuilderConfig):
    language: Optional[str] = None
    with_audio: bool = True
    decode_audio: bool = True
    sampling_rate: Optional[int] = None
    mono: bool = True


class JamAltDataset(datasets.GeneratorBasedBuilder):
    _DESCRIPTION

    VERSION = datasets.Version(_VERSION)
    BUILDER_CONFIG_CLASS = JamAltBuilderConfig
    BUILDER_CONFIGS = [JamAltBuilderConfig("all")] + [
        JamAltBuilderConfig(lang, language=lang)
        for lang in _LANGUAGE_NAME_TO_CODE.values()
    ]
    DEFAULT_CONFIG_NAME = "all"

    def _info(self):
        feat_dict = {
            "name": datasets.Value("string"),
            "text": datasets.Value("string"),
            "language": datasets.Value("string"),
            "license_type": datasets.Value("string"),
        }
        if self.config.with_audio:
            feat_dict["audio"] = datasets.Audio(
                decode=self.config.decode_audio,
                sampling_rate=self.config.sampling_rate,
                mono=self.config.mono,
            )

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(feat_dict),
            supervised_keys=("audio", "text") if "audio" in feat_dict else None,
            homepage=_HOMEPAGE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        metadata_path = dl_manager.download(_METADATA_FILENAME)

        audio_paths, text_paths, metadata = [], [], []
        with open(metadata_path, encoding="utf-8") as f:
            for row in csv.DictReader(f):
                if (
                    self.config.language is None
                    or _LANGUAGE_NAME_TO_CODE[row["Language"]] == self.config.language
                ):
                    audio_paths.append("audio/" + row["Filepath"])
                    text_paths.append(
                        "lyrics/" + os.path.splitext(row["Filepath"])[0] + ".txt"
                    )
                    metadata.append(row)

        text_paths = dl_manager.download(text_paths)
        audio_paths = (
            dl_manager.download(audio_paths) if self.config.with_audio else None
        )

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs=dict(
                    text_paths=text_paths,
                    audio_paths=audio_paths,
                    metadata=metadata,
                ),
            ),
        ]

    def _generate_examples(self, text_paths, audio_paths, metadata):
        if audio_paths is None:
            audio_paths = [None] * len(text_paths)
        for text_path, audio_path, meta in zip(text_paths, audio_paths, metadata):
            name = os.path.splitext(meta["Filepath"])[0]
            with open(text_path, encoding="utf-8") as text_f:
                record = {
                    "name": name,
                    "text": text_f.read(),
                    "language": _LANGUAGE_NAME_TO_CODE[meta["Language"]],
                    "license_type": meta["LicenseType"],
                }
            if audio_path is not None:
                record["audio"] = audio_path
            yield name, record