my_stt_dataset.py

import os
import csv
import datasets
from datasets import Audio, BuilderConfig

# Konfiguratsiya sinfi: til qisqartmasi va ma'lumotlar joylashgan papkani belgilaydi.
class STTConfig(BuilderConfig):
    def __init__(self, language_abbr, data_dir, **kwargs):
        """
        Args:
            language_abbr (str): Masalan, "uz".
            data_dir (str): Dataset joylashgan asosiy papka, masalan "Dataset_STT".
            **kwargs: Qolgan parametrlar.
        """
        super().__init__(**kwargs)
        self.language_abbr = language_abbr
        self.data_dir = data_dir

# Dataset yuklash skripti
class MySTTDataset(datasets.GeneratorBasedBuilder):
    """
    Uzbek STT dataset yuklash skripti:
      - Audio fayllar .tar arxiv ichida saqlangan.
      - Transkripsiya ma'lumotlari TSV faylda joylashgan.
      - Streaming rejimida, tar fayllar dl_manager.iter_archive() orqali o‘qiladi.
      - "audio" ustuni Audio() tipida aniqlangan, ya'ni qiymat dictionary shaklida:
            {"path": <tar ichidagi fayl nomi>, "bytes": <audio baytlari>}
        bo‘lishi kerak, shunda Dataset Viewer "play" tugmasini ko‘rsatadi.
    """
    VERSION = datasets.Version("1.0.0")

    BUILDER_CONFIGS = [
        STTConfig(
            name="uz",
            version=datasets.Version("1.0.0"),
            description="Uzbek subset of the STT dataset",
            language_abbr="uz",
            data_dir="Dataset_STT",  # Asosiy papka nomi
        )
    ]
    DEFAULT_CONFIG_NAME = "uz"

    def _info(self):
        """
        Dataset ustunlarini aniqlaydi.
        "audio" ustuni Audio(sampling_rate=None) tipida berilgan, shuning uchun
        audio fayllar avtomatik dekodlanadi va resample qilinadi.
        """
        return datasets.DatasetInfo(
            description=(
                "Uzbek STT dataset: audio fayllar tar arxivida saqlangan va "
                "transcriptions esa TSV faylda mavjud. Streaming rejimi bilan tar "
                "arxivdan audio fayllar o'qiladi."
            ),
            features=datasets.Features({
                "id": datasets.Value("string"),
                "audio": Audio(sampling_rate=None),
                "sentence": datasets.Value("string"),
                "duration": datasets.Value("float"),
                "age": datasets.Value("string"),
                "gender": datasets.Value("string"),
                "accents": datasets.Value("string"),
                "locale": datasets.Value("string"),
            }),
            supervised_keys=None,
            version=self.VERSION,
        )

    def _split_generators(self, dl_manager):
        """
        Har bir split uchun: tar arxiv va mos TSV fayllarining yo'llari aniqlanadi.
        Tar arxivlardan streaming rejimida o'qish uchun dl_manager.iter_archive() dan foydalanamiz.
        """
        config = self.config
        base_dir = config.data_dir  # Masalan: "Dataset_STT"
        lang = config.language_abbr   # Masalan: "uz"

        # Tar arxiv fayllari (extract qilinmaydi, balki iter_archive orqali o'qiladi)
        train_tar = os.path.join(base_dir, "audio", lang, "train.tar")
        test_tar = os.path.join(base_dir, "audio", lang, "test.tar")
        val_tar = os.path.join(base_dir, "audio", lang, "validation.tar")

        train_audio_files = dl_manager.iter_archive(train_tar)
        test_audio_files = dl_manager.iter_archive(test_tar)
        val_audio_files = dl_manager.iter_archive(val_tar)

        # TSV fayllar yo'li
        train_tsv = os.path.join(base_dir, "transcript", lang, "train.tsv")
        test_tsv = os.path.join(base_dir, "transcript", lang, "test.tsv")
        val_tsv = os.path.join(base_dir, "transcript", lang, "validation.tsv")

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"audio_files": train_audio_files, "tsv_path": train_tsv},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"audio_files": test_audio_files, "tsv_path": test_tsv},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"audio_files": val_audio_files, "tsv_path": val_tsv},
            ),
        ]

    def _generate_examples(self, audio_files, tsv_path):
        """
        TSV faylini qatorma-qator o'qiydi va metadata lug'atini tuzadi.
        So'ng, tar arxividan kelayotgan audio fayllarni (streaming iteratori orqali)
        mos metadata bilan birlashtiradi.
        
        Har bir audio ustuni qiymati quyidagicha shakllantiriladi:
            {"path": <tar ichidagi fayl nomi>, "bytes": <audio fayl baytlari>}
        Bu shakl Dataset Viewer tomonidan Audio() sifatida aniqlanadi.
        """
        # TSV faylidan metadata lug'atini tuzamiz: kalit – fayl nomi (masalan, "ID.mp3")
        metadata = {}
        with open(tsv_path, "r", encoding="utf-8") as f:
            reader = csv.DictReader(f, delimiter="\t")
            for row in reader:
                filename = row["id"] + ".mp3"
                metadata[filename] = row

        # Tar arxivdan streaming iterator orqali o'qilgan fayllar
        for idx, (file_path, file_obj) in enumerate(audio_files):
            # file_path: tar arxiv ichidagi nisbiy yo'l (masalan, "009f0d56-c7db-4de3-bd3e-92a37d6f0cb9.mp3")
            if file_path in metadata:
                row = metadata[file_path]
                audio_bytes = file_obj.read()
                yield idx, {
                    "id": row["id"],
                    "audio": {"path": file_path, "bytes": audio_bytes},
                    "sentence": row.get("sentence", ""),
                    "duration": float(row.get("duration", 0.0)),
                    "age": row.get("age", ""),
                    "gender": row.get("gender", ""),
                    "accents": row.get("accents", ""),
                    "locale": row.get("locale", ""),
                }