my_stt_dataset.py

import os
import tarfile
import csv

import datasets
from datasets import Audio


class MySTTDatasetConfig(datasets.BuilderConfig):
    """Config klass (kerak bo'lsa turli versiya yoki param qo'yish uchun)."""


class MySTTDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        MySTTDatasetConfig(name="default", version=datasets.Version("1.0.0")),
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description="STT dataset .tar ichida audio, .tsv ichida transkript",
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "audio": Audio(sampling_rate=None),  # Audio turi
                    "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,
        )

    def _split_generators(self, dl_manager):
        """
          dl_manager.download_and_extract() bilan remote fayllarni yuklaysiz
          yoki local path bersangiz bo'ladi.
          Masalan: 
            https://huggingface.co/datasets/Elyordev/new_dataset_stt/resolve/main/Dataset_STT/audio/uz/train/train.tar
            https://huggingface.co/datasets/Elyordev/new_dataset_stt/resolve/main/Dataset_STT/transcript/uz/train/train.tsv
        """
        train_tar = dl_manager.download_and_extract("URL_train_tar")  # masalan
        train_tsv = dl_manager.download("URL_train_tsv")

        val_tar = dl_manager.download_and_extract("URL_val_tar")
        val_tsv = dl_manager.download("URL_val_tsv")

        test_tar = dl_manager.download_and_extract("URL_test_tar")
        test_tsv = dl_manager.download("URL_test_tsv")

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "archive_path": train_tar,
                    "tsv_path": train_tsv,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "archive_path": val_tar,
                    "tsv_path": val_tsv,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "archive_path": test_tar,
                    "tsv_path": test_tsv,
                },
            ),
        ]

    def _generate_examples(self, archive_path, tsv_path):
        """
         Har bir tar faylni ochib, TSV dagi id + sentence boshqariladi.
        """
        # `archive_path` bu dl_manager.download_and_extract qilgan manzil
        # lekin e’tibor bering: .extract automatik ravishda chiqarib yuboradi.
        # Agar `.tar` ochilmasin desangiz, download() deyish kifoya, 
        # lekin streaming qilish uchun biroz boshqacha yo'l tutish kerak.

        # Keling, bu yerda tar fayl allaqachon extract qilingan deb faraz qilamiz:
        #  train.tar -> train/ (ichida mp3 fayllar paydo bo'lgan bo'ladi).
        #  Agar real .tar ichidan "on-the-fly" o'qimoqchi bo'lsak, 
        #  dl_manager.download() + tarfile.open(...) da ishlash lozim.

        # Shunchaki misol tariqasida:
        audio_base_path = archive_path  # extract bo'lgach papka manzili
        # Ehtimol, audio_base_path = os.path.join(archive_path, "train") bo'lishi ham mumkin
        # chunki tar ichida "train/" deb nomlangan ichki papka bo'lishi mumkin.

        with open(tsv_path, "r", encoding="utf-8") as f:
            reader = csv.DictReader(f, delimiter="\t")
            for n, row in enumerate(reader):
                audio_id = row["id"]  # masalan '009f0d56-c7db-4de3-bd3e-92a37d6f0cb9'
                audio_file = audio_id + ".mp3"
                
                # to'liq path
                audio_path = os.path.join(audio_base_path, audio_file)

                # Agar audio fayl topilsa:
                if os.path.isfile(audio_path):
                    yield n, {
                        "id": audio_id,
                        "audio": audio_path,  # Audio typeda faqat path bersak, HF o'zi o'qiydi
                        "sentence": row["sentence"],
                        "duration": float(row["duration"]),
                        "age": row["age"],
                        "gender": row["gender"],
                        "accents": row["accents"],
                        "locale": row["locale"],
                    }
                else:
                    # Agar mp3 topilmasa, o'ziz xato signal qilishingiz mumkin
                    # yoki continue qilishingiz mumkin
                    pass