dataset_stt.py

import csv
import os
import tarfile

import datasets
from tqdm import tqdm

_DESCRIPTION = """\
This dataset is designed for speech-to-text (STT) tasks. It contains audio files stored as tar archives along with their corresponding transcript files in TSV format. The data is for the Uzbek language.
"""

_CITATION = """\
@misc{dataset_stt2025,
  title={Dataset_STT},
  author={Your Name},
  year={2025}
}
"""

class DatasetSTT(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")
    
    def _info(self):
        features = datasets.Features({
            "id": datasets.Value("string"),
            "audio": datasets.Audio(sampling_rate=16000),  # Agar kerak bo'lsa, sampling_rate ni moslashtiring
            "sentence": datasets.Value("string"),
            "duration": datasets.Value("float"),
            "age": datasets.Value("string"),
            "gender": datasets.Value("string"),
            "accents": datasets.Value("string"),
            "locale": datasets.Value("string")
        })
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            supervised_keys=None,
            homepage="https://huggingface.co/datasets/Elyordev/Dataset_STT",
            citation=_CITATION,
        )
    
    def _split_generators(self, dl_manager):
        """
        _split_generators da har bir split uchun kerakli fayllarni belgilaymiz.
        Biz quyidagi splitlarni qo'llaymiz: TRAIN, TEST va VALIDATION.
        Data_files argumenti orqali audio arxiv va transcript TSV fayllarini olamiz.
        """
        data_files = self.config.data_files
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "audio_archive": data_files["train"]["audio"],
                    "transcript_file": data_files["train"]["transcript"],
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "audio_archive": data_files["test"]["audio"],
                    "transcript_file": data_files["test"]["transcript"],
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "audio_archive": data_files["validation"]["audio"],
                    "transcript_file": data_files["validation"]["transcript"],
                },
            ),
        ]
    
    def _generate_examples(self, audio_archive, transcript_file):
        """
        Transcript TSV faylini o'qib, har bir yozuv uchun:
          - Tar arxivni ochamiz va audio fayllarni indekslaymiz.
          - Transcript faylida ko'rsatilgan "path" ustuni orqali mos audio faylni topamiz.
          - Audio faylni butun baytlar shaklida o'qib, audio maydoni sifatida qaytaramiz.
        """
        # Tar arxivni ochamiz
        with tarfile.open(audio_archive, "r:*") as tar:
            # Arxiv ichidagi barcha fayllarni (fayl nomi -> tarinfo) indekslaymiz
            tar_index = {os.path.basename(member.name): member for member in tar.getmembers() if member.isfile()}
            
            # Transcript TSV faylini ochamiz (UTF-8 kodlashda)
            with open(transcript_file, "r", encoding="utf-8") as f:
                reader = csv.DictReader(f, delimiter="\t")
                for row in tqdm(reader, desc="Processing transcripts"):
                    file_name = row["path"]  # Masalan: "2cd08f62-aa25-4f5e-bb73-40cfc19a215e.mp3"
                    if file_name not in tar_index:
                        print(f"Warning: {file_name} not found in {audio_archive}")
                        continue
                    
                    audio_member = tar.extractfile(tar_index[file_name])
                    if audio_member is None:
                        print(f"Warning: Could not extract {file_name}")
                        continue
                    
                    audio_bytes = audio_member.read()
                    
                    yield row["id"], {
                        "id": row["id"],
                        "audio": {"path": file_name, "bytes": audio_bytes},
                        "sentence": row["sentence"],
                        "duration": float(row["duration"]) if row["duration"] else 0.0,
                        "age": row["age"],
                        "gender": row["gender"],
                        "accents": row["accents"],
                        "locale": row["locale"],
                    }