my_stt_dataset.py

import os
import csv
import datasets
from datasets import Audio


class MySTTDataset(datasets.GeneratorBasedBuilder):
    """
    Common Voice uslubidagi minimal dataset skript:
      - 3 ta tar fayl (train/test/validation)
      - Har bir tar fayl ichida .mp3 audio
      - Har bir split'ga mos TSV fayl (train.tsv, test.tsv, validation.tsv)
      - Audio ustuni -> HF Viewer da "play" tugmasi
    """
    VERSION = datasets.Version("1.0.0")

    # Agar ko'p config bo'lmasa, bu qismni soddalashtiramiz.
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="uz",
            version=VERSION,
            description="STT dataset for Uzbek language (example).",
        )
    ]

    DEFAULT_CONFIG_NAME = "uz"

    def _info(self):
        """
        Bu yerda datasetning xususiyatlari (features) e'lon qilinadi.
        'audio' ustuni Audio() turida bo'lsa, viewer pleyer ko'rsatadi.
        """
        return datasets.DatasetInfo(
            description="Uzbek STT dataset: audio in .tar, transcriptions in .tsv.",
            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 va tsv fayllar yo'lini belgilab, 
        dl_manager orqali yuklab/extract qildirib, so'ng _generate_examples() ga beramiz.
        """
        # local path misoli (reposingizda bo'lsa). 
        # Agar huggingface.co'dan yuklamoqchi bo'lsangiz, URL qilishingiz mumkin
        train_tar = "Dataset_STT/audio/uz/train.tar"
        train_tsv = "Dataset_STT/transcript/uz/train.tsv"

        test_tar = "Dataset_STT/audio/uz/test.tar"
        test_tsv = "Dataset_STT/transcript/uz/test.tsv"

        val_tar = "Dataset_STT/audio/uz/validation.tar"
        val_tsv = "Dataset_STT/transcript/uz/validation.tsv"

        # Bu fayllarni download+extract (yoki local bo'lsa, faqat extract) qilamiz:
        # Eslatma: agar localda bo'lsayu, dl_manager `is_local=True` deb topishi mumkin,
        # ammo baribir .extract ishlaydi.

        train_tar_extracted = dl_manager.extract(train_tar)
        test_tar_extracted = dl_manager.extract(test_tar)
        val_tar_extracted = dl_manager.extract(val_tar)

        # Har bir splitted datasetga mos "SplitGenerator" qaytaramiz
        # "gen_kwargs" -> _generate_examples() ga paramlar
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "archive_dir": train_tar_extracted,  # tar fayl ochilib yoyilgan papka
                    "tsv_path": train_tsv,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "archive_dir": test_tar_extracted,
                    "tsv_path": test_tsv,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "archive_dir": val_tar_extracted,
                    "tsv_path": val_tsv,
                },
            ),
        ]

    def _generate_examples(self, archive_dir, tsv_path):
        """
        Ushbu metod har bir split uchun audio+transkript juftliklarini geneate qiladi.
          - 'archive_dir' papkada .tar dan ochilgan .mp3 fayllar mavjud.
          - 'tsv_path' faylini qatorma-qator o'qib, 'id' -> "id.mp3" yo'lini izlaymiz.
        """
        # TSV ni o'qiymiz:
        with open(tsv_path, "r", encoding="utf-8") as f:
            reader = csv.DictReader(f, delimiter="\t")
            for idx, row in enumerate(reader):
                # tsv da shunaqa ustunlar bo'lishi kutiladi:
                # id, sentence, duration, age, gender, accents, locale
                audio_id = row["id"]
                mp3_file = audio_id + ".mp3"
                mp3_path = os.path.join(archive_dir, mp3_file)

                # Agar audio fayl exist bo'lsa:
                if os.path.isfile(mp3_path):
                    yield idx, {
                        "id": audio_id,
                        "audio": mp3_path,  # Audio() -> pleyer
                        "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", ""),
                    }
                else:
                    # Audio topilmasa, skip (yoki exception)
                    continue