create_release.py

# This scripts generates a self-contained dataset relase that can be uploaded to the HF-Hub or distributed as an archive file

# Apart from pip-installable packages this file should be self-contained for reproducibility
import json
import re
import shutil
from pathlib import Path
import jinja2
import pandas as pd
import uroman as ur
from unidecode import unidecode
from lnn.utils import load_audio
from tqdm.auto import tqdm

# Define some variables
NAME = "SpeechTaxi"
VERSION = "0.0.1"
HOMEPAGE = "github.com/LennartKeller"
CITATION = """
@misc{keller2024speechtaximultilingualsemanticspeech,
      title={SpeechTaxi: On Multilingual Semantic Speech Classification}, 
      author={Lennart Keller and Goran Glavaš},
      year={2024},
      eprint={2409.06372},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.06372}, 
}
""".strip()
DESCRIPTION = "Multlingual Semantic Speech Classification"
SAVE_DIR = Path("/hdd/lennart/datasets")

# Define some arguments and 'constants'
# TODO think about ways to avoid having to do this
REWRITE_PATHS = (
    "/Volumes/SanDiskData/bible-audio-scraper-data/processed",
    "/hdd/lennart/bible-audio-scraper-data/processed",
)
REWRITE_MASS_PATHS = ("/Volumes/SanDiskData/", "/hdd/lennart/")
ALIGNED_TABLES_DIR = Path("aligned_tables")
INCLUDE_MASS_LANGUAGES = False
MASS_LANGUAGES = ("eng", "rus", "spa", "ron", "eus", "fra")
SKIP_LANGUAGES = ("ron", "eus", "spa", "hau", "bdv", "nep", "ben")
SCRIPT_TEMPLATE = Path("SpeechTaxi.py.jinja")

if not INCLUDE_MASS_LANGUAGES:
    SKIP_LANGUAGES += MASS_LANGUAGES

def rewrite_path(p: str | Path) -> str | Path:
    if REWRITE_PATHS:
        if isinstance(p, str):
            p = p.replace(*REWRITE_PATHS)
        p = Path(str(p).replace(*REWRITE_PATHS))
    return p


def rewrite_mass_path(p: str | Path) -> str | Path:
    if REWRITE_MASS_PATHS:
        if isinstance(p, str):
            p = p.replace(*REWRITE_MASS_PATHS)
        p = Path(str(p).replace(*REWRITE_MASS_PATHS))
    return p


# Define a 'filter_instances' function that holds the selection logic
# for selecting instances that shold end up in the final dataset
def filter_instances(
    df: pd.DataFrame, table_file_path: str | Path = None
) -> pd.DataFrame:
    df = df.copy()  # Avoid warnings
    # Some mass tables do not have them...
    if "alignment_score" not in df.columns:
        df["alignment_score"] = 1.0
    df = df.query("score == 1.0 and alignment_score >= 0.8")
    return df


def render_script_template(**kwargs):
    templateLoader = jinja2.FileSystemLoader(
        searchpath=SCRIPT_TEMPLATE.parent.as_posix()
    )
    templateEnv = jinja2.Environment(loader=templateLoader)
    template = templateEnv.get_template(SCRIPT_TEMPLATE.name)
    return template.render(**kwargs)


# BEGIN LIB:
# Define libs that are used to create the dataset.
# Do not change stuff here unless you really want to.

BOOK_ABBREVIATIONS = {
    "Matthew": "MAT",
    "Mark": "MRK",
    "Luke": "LUK",
    "John": "JHN",
    "Acts": "ACT",
    "Romans": "ROM",
    "Corinthians": "COR",  # Can be used as "1COR" or "2COR" based on context
    "Galatians": "GAL",
    "Ephesians": "EPH",
    "Philippians": "PHP",
    "Colossians": "COL",
    "Thess": "THS",  # Can be used as "1THS" or "2THS" based on context
    "Timothy": "TIM",  # Can be used as "1TIM" or "2TIM" based on context
    "Titus": "TIT",
    "Philemon": "PHM",
    "Hebrews": "HEB",
    "James": "JAS",
    "Peter": "PET",  # Can be used as "1PET" or "2PET" based on context
    "Jude": "JUD",
    "Revelation": "REV",
}
BOOK_ABBREVIATIONS_FR = {
    "Matthieu": "MAT",
    "Marc": "MRK",
    "Luc": "LUC",
    "Jean": "JHN",
    "Actes": "ACT",
    "Romains": "ROM",
    "Corinthiens": "COR",  # Can be used as "1COR" or "2COR" based on context
    "Galates": "GAL",
    "Éphésiens": "EPH",
    "Philippiens": "PHP",
    "Colossiens": "COL",
    "Thessaloniciens": "THS",  # Can be used as "1THS" or "2THS" based on context
    "Timothée": "TIM",  # Can be used as "1TIM" or "2TIM" based on context
    "Tite": "TIT",
    "Philémon": "PHM",
    "Hébreux": "HEB",
    "Jacques": "JAC",
    "Pierre": "PIE",  # Can be used as "1PIE" or "2PIE" based on context
    "Jude": "JUD",
    "Apocalypse": "APO",
}

BOOK_ABBREVIATIONS |= {
    unidecode(k): unidecode(v) for k, v in BOOK_ABBREVIATIONS_FR.items()
}

MASS_LANGUAGE_IDS = [
    "FRNTLSN2DA",
    "ENGESVN1DA",
    "RUSS76N2DA",
    "SPNBDAN1DA",
    "EUSEABN1DA",
    "RONDCVN1DA",
]


def parse_reference_from_mass_file(filename: str | Path, language_id: str = ""):
    filename = Path(filename).name
    book_num, rest = filename.split("___", 1)
    book_num = int(book_num.replace("B", ""))
    chapter_num, rest = rest.split("_", 1)
    chapter_num = int(chapter_num)
    if language_id:
        rest = rest.replace(language_id, "")
    else:
        # rest = "_".join(p for p in rest.split("_") if not p.endswith("DA"))
        for language_id in MASS_LANGUAGE_IDS:
            rest = rest.replace(language_id, "")
    rest = rest.replace("_one_channel.wav", "")
    book_ref, verse_ref = re.split(r"_+", rest, 1)
    book_num = ""
    if book_ref[0].isdigit():
        book_num = book_ref[0]
        book_ref = book_ref[1:]
    try:
        book_ref = BOOK_ABBREVIATIONS[book_ref]
    except KeyError as e:
        print(book_ref)
        print(filename)
        raise e
    if book_num:
        book_ref = book_num + book_ref
    verse_num = int(verse_ref.split("_")[-1])
    complete_ref = f"{book_ref}.{chapter_num}.{verse_num}"
    return complete_ref


def copy_self(dst: str | Path) -> None:
    dst = Path(dst)
    path_to_self = Path(__file__)
    dst = dst / path_to_self.name
    dst.write_text(path_to_self.read_text())


def copy_bible_scrape_audio(file: str | Path, root: str | Path) -> Path:
    file, root = map(Path, (file, root))
    copied_file = (
        root / file.parts[-4] / file.parts[-3] / file.parts[-2] / file.parts[-1]
    )
    copied_file.parent.mkdir(exist_ok=True, parents=True)
    shutil.copy(file, copied_file)
    return copied_file


def copy_mass_audio(file: str | Path, root: str | Path) -> Path:
    file, root = map(Path, (file, root))
    copied_file = root / file.parts[-2] / file.parts[-1]
    copied_file.parent.mkdir(exist_ok=True, parents=True)
    shutil.copy(file, copied_file)
    return copied_file


uroman = ur.Uroman()


def read_mass_instance(audio_file: str | Path) -> dict:
    audio_file = Path(audio_file)
    audio_file = rewrite_path(audio_file)
    if not audio_file.exists():
        raise ValueError(f"Audio file {audio_file} does not exist.")
    verse_ref = parse_reference_from_mass_file(audio_file)
    transcription_file = (
        audio_file.parent / f"{audio_file.stem.removesuffix('_one_channel')}.txt"
    )
    transcription = transcription_file.read_text()
    machine_generated_transcriptions = {}
    machine_generated_transcription_files = audio_file.parent.glob(
        f"{audio_file.stem}-*-transcript.txt"
    )
    for machine_generated_transcription_file in machine_generated_transcription_files:
        # Get transcription key
        key = machine_generated_transcription_file.stem.split("_one_channel-")[-1]
        key = key.removesuffix("-transcript")
        # print(audio_file, "|", machine_generated_transcription_file, "|", key)
        machine_generated_transcriptions[f"transcription_{key}"] = (
            machine_generated_transcription_file.read_text()
        )
    data = {
        "verse_ref": verse_ref,
        "transcription": transcription,
        "transcription_romanized": uroman.romanize_string(transcription),
    } | machine_generated_transcriptions
    return data


def read_bible_scrape_instance(audio_file: str | Path) -> dict:
    """
    Given the file to a verse this function loads all available metadata from disk.
    And ensure that the audio files exists...
    """
    audio_file = Path(audio_file)
    audio_file = rewrite_path(audio_file)

    if not audio_file.exists():
        raise ValueError(f"Audio file {audio_file} does not exist.")
    # 1. Load all avaiable transcriptions
    transcription_file = audio_file.with_suffix(".json")
    transcriptions = json.loads(transcription_file.read_text())
    # Filter to only texts
    try:
        verse_ref = transcriptions.pop("verse_ref")
        transcriptions = {
            bible_id: content["content"] for bible_id, content in transcriptions.items()
        }
    except Exception as e:
        print(transcriptions)
        print(e)
    # 2. Look for audio alignement scores
    alignment_file = audio_file.parent / f"{audio_file.stem}-alignment-scores.json"
    alignment_scores = json.loads(alignment_file.read_text())

    # 3. Select the transcription with best audio alignment score
    best_bible_id = max(alignment_scores.items(), key=lambda x: x[-1])[0]
    transcription = transcriptions[str(best_bible_id)]
    # 4. Load all available machine generated transcriptions
    machine_generated_transcriptions = {}
    machine_generated_transcription_files = audio_file.parent.glob(
        f"{audio_file.stem}-*.txt"
    )
    for machine_generated_transcription_file in machine_generated_transcription_files:
        # Get transcription key

        key = (
            machine_generated_transcription_file.stem.removeprefix(audio_file.stem)
            .removesuffix("-transcript")
            .strip("-")
        )
        # if key.endswith("mms"):
        #     continue
        # print(audio_file, "|", machine_generated_transcription_file, "|", key)
        machine_generated_transcriptions[f"transcription_{key}"] = (
            machine_generated_transcription_file.read_text()
        )

    data = {
        "verse_ref": verse_ref,
        "transcription": transcription,
        "transcription_romanized": uroman.romanize_string(transcription),
    } | machine_generated_transcriptions
    return data


def is_audio_valid(audio_path: str | Path):
    audio_path = Path(audio_path)
    if not audio_path.exists():
        return False
    try:
        wv, sr = load_audio(audio_path, return_tensor="torch")
        wv = wv[0].reshape(-1)
    except Exception:
        return False
    # We need *at least* one second of audio
    return wv.numel() >= sr


# Main logic
if __name__ == "__main__":
    dataset_root = SAVE_DIR / NAME
    # Setup the output directory (-ies)
    if dataset_root.exists():
        raise ValueError(
            f"Dataset-Root-Directory {dataset_root} already exists, please remove manually to continue!"
        )
    print(f"Creating dataset release in {dataset_root}")
    dataset_root.mkdir(parents=True)
    data_dir = dataset_root / "data"
    data_dir.mkdir(parents=True)
    table_dir = dataset_root / "tables"
    table_dir.mkdir(parents=True)

    # Copy this script to dataset dir for reproducibility
    copy_self(dataset_root)

    # Create language datasets and copy the data to the dataset directory
    # Load and filter alignment tables
    language_tables = sorted(
        [f for f in ALIGNED_TABLES_DIR.glob("*.tsv") if f.stem not in SKIP_LANGUAGES]
    )

    # Write code file
    success_languages = []
    pbar = tqdm(language_tables, desc="Creating dataset...")
    invalid_counter = 0
    for alignment_table in pbar:
        df = pd.read_table(alignment_table)
        try:
            df = filter_instances(df, alignment_table)
        except Exception as e:
            print(alignment_table)
            raise e
        if not len(df):
            continue
        success_languages.append(alignment_table)
        final_table_data = []
        for _, row in df.iterrows():
            if alignment_table.stem not in MASS_LANGUAGES:
                audio_file = rewrite_path(Path(row["audio_path"]))
                if not is_audio_valid(audio_file):
                    invalid_counter += 1
                    continue
                copied_audio_file = copy_bible_scrape_audio(audio_file, root=data_dir)
                english_text = row["en_text"]
                try:
                    data = read_bible_scrape_instance(audio_file)
                except Exception as e:
                    print("Error reading file", audio_file)
                    print(e)
                    print("_" * 30)
                    continue
                # Files that can't be read have zero alignment scores anyways so in most cases this isn't an issue
                data = (
                    {"verse_ref": data.pop("verse_ref")}
                    | {"text_en": english_text}
                    | {"split": row["split"], "label": row["label"]}
                    | data
                    | {"audio": copied_audio_file.relative_to(data_dir).as_posix()}
                )
                final_table_data.append(data)
            elif INCLUDE_MASS_LANGUAGES:
                audio_file = rewrite_mass_path(Path(row["audio_path"]))
                if not is_audio_valid(audio_file):
                    invalid_counter += 1
                    continue
                copied_audio_file = copy_mass_audio(audio_file, root=data_dir)
                try:
                    data = read_mass_instance(audio_file)
                except Exception as e:
                    print("Error reading file", audio_file)
                    print(e)
                    print("_" * 30)
                    continue
                # Files that can't be read have zero alignment scores anyways so in most cases this isn't an issue
                data = data | {
                    "split": row["split"],
                    "label": row["label"],
                    "text_en": row["en_text"],
                    "audio": copied_audio_file.relative_to(data_dir).as_posix(),
                }
                final_table_data.append(data)
            pbar.set_description(
                f"{str(alignment_table)} | Total invalid: {invalid_counter}"
            )

        final_table = pd.DataFrame.from_records(final_table_data)
        final_table.to_csv(table_dir / alignment_table.name, index=False, sep="\t")

    # Now render dataset loading script and write to dataset dir
    code_file = dataset_root / f"{NAME}.py"
    code_file_content = render_script_template(
        languages=[f.stem for f in success_languages],
        version=VERSION,
        homepage=HOMEPAGE,
        citation=CITATION,
    )
    code_file.write_text(code_file_content)

    # Zip data and tables dir
    shutil.make_archive(dataset_root / "tables", "zip", table_dir)
    shutil.rmtree(table_dir)
    shutil.make_archive(dataset_root / "data", "zip", data_dir)
    shutil.rmtree(data_dir)