prepare_dataset.py

import dataclasses
from pathlib import Path
import shutil
import csv
from typing import Any, Optional

import srt
import soundfile as sf


@dataclasses.dataclass
class Subtitle:
    start: float
    end: float
    content: str

    @property
    def duration(self):
        return self.end - self.start

    @classmethod
    def from_sub(cls, sub: srt.Subtitle):
        return cls(
            start=sub.start.total_seconds(),
            end=sub.end.total_seconds(),
            content=sub.content
        )

    @classmethod
    def merge(cls, sub1: 'Subtitle', sub2: 'Subtitle'):
        return cls(
            start=sub1.start,
            end=sub2.end,
            content=f"{sub1.content.strip()} {sub2.content.strip()}"
        )


def merge_subs(subs: list[Subtitle], max_sample_duration_s: float) -> list[Subtitle]:
    merged_subs: list[Subtitle] = []
    previous_sub: Subtitle = subs[0]
    for sub in subs[1:]:
        # there is a short silence between subs or we reach max duration
        if previous_sub.end + 0.2 < sub.start or previous_sub.duration + sub.duration > max_sample_duration_s:
            merged_subs.append(previous_sub)
            previous_sub = sub
        elif previous_sub and previous_sub.end + 0.2 >= sub.start:
            previous_sub = Subtitle.merge(previous_sub, sub)
        else:
            raise ValueError("Subtitles are not in order")
    merged_subs.append(previous_sub)
    return merged_subs


original_audios_path = Path("darija-test-folder")
dataset_path = Path("audio_dataset/data/test")
max_sample_duration_s = 30

if dataset_path.exists():
    shutil.rmtree(dataset_path)

dataset_path.mkdir(parents=True)

file_to_subs: dict[Path, list[Subtitle]] = {}
for file in original_audios_path.iterdir():
    if file.suffix == ".wav":
        with open(file.parent / f"{file.stem}.srt", "r") as f:
            subs = srt.parse(f.read())
        file_to_subs[file] = merge_subs([Subtitle.from_sub(sub) for sub in subs],
                                        max_sample_duration_s)

columns = ["file_name", "transcription", "sample_id", "start_timestamp", "end_timestamp", "audio_name"]
csv_lines: list[dict[str, Any]] = []

for file in file_to_subs:
    audio, sr = sf.read(file)
    for sub in file_to_subs[file]:
        file_name = f"{file.stem}-{sub.start:.2f}-{sub.end:.2f}.wav"
        audio_cut = audio[int(round(sub.start * sr)): int(round(sub.end * sr))]
        sf.write(dataset_path / file_name, audio_cut, sr)
        csv_lines.append({
            "file_name": Path("data") / "test" / file_name,
            "transcription": sub.content,
            "sample_id": Path(file_name).stem,
            "start_timestamp": sub.start,
            "end_timestamp": sub.end,
            "audio_name": file.stem
        })

with (dataset_path.parent / "metadata.csv").open('w', encoding="utf-8") as f:
    writer = csv.DictWriter(f, fieldnames=columns)
    writer.writeheader()
    writer.writerows(csv_lines)