my_stt_dataset.py

import os
import csv
import datasets
from datasets import Audio

# Har xil config - 'sample' va 'full'
class MySTTDatasetConfig(datasets.BuilderConfig):
    def __init__(self, limit=None, **kwargs):
        """
        limit : int yoki None
          Har bir splitdan qancha qatorni o'qish.
          None bo'lsa, cheklanmagan.
        """
        super().__init__(**kwargs)
        self.limit = limit


class MySTTDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        MySTTDatasetConfig(
            name="sample",
            version=datasets.Version("1.0.0"),
            description="Faqat har bir splitdan 10k qator ko'rsatish uchun",
            limit=10_000,  # masalan 10 ming
        ),
        MySTTDatasetConfig(
            name="full",
            version=datasets.Version("1.0.0"),
            description="Hech qanday cheklovsiz to'liq dataset",
            limit=None,
        ),
    ]
    
    DEFAULT_CONFIG_NAME = "sample"
    
    def _info(self):
        return datasets.DatasetInfo(
            description="Speech-to-text dataset (tar ichida audio, tsvda transkript).",
            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,
        )
    
    def _split_generators(self, dl_manager):
        # TODO: train.tar, test.tar, validation.tar + tegishli TSV link yoki local path
        # Hozircha misol tariqasida local path'lar ko'rsatamiz
        train_tar = "Dataset_STT/audio/uz/train/train.tar"
        train_tsv = "Dataset_STT/transcript/uz/train/train.tsv"
        
        val_tar = "Dataset_STT/audio/uz/validation/validation.tar"
        val_tsv = "Dataset_STT/transcript/uz/validation/validation.tsv"
        
        test_tar = "Dataset_STT/audio/uz/test/test.tar"
        test_tsv = "Dataset_STT/transcript/uz/test/test.tsv"
        
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "tar_path": train_tar,
                    "tsv_path": train_tsv,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "tar_path": val_tar,
                    "tsv_path": val_tsv,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "tar_path": test_tar,
                    "tsv_path": test_tsv,
                },
            ),
        ]
    
    def _generate_examples(self, tar_path, tsv_path):
        """
        limit=10_000 bo'lsa, har bir splitdan 10 mingtagina qator qaytaradi.
        Agar limit=None bo'lsa, hamma qatorni o'qiydi.
        """
        limit = self.config.limit
        
        # Tar ichidagi mp3 fayllarni avval extract qilasiz yoki on-the-fly o'qiysiz
        # Eslatma: HF Viewer uchun eng osoni audio papkaga ochib qo'yish yoki
        # `dl_manager.download_and_extract(...)` ishlatishdir.
        # Bu yerda misol tariqasida tar ni ochib, audio fayllarni papkaga yoyildi deb faraz qilamiz:
        
        # Masalan audio papka: "Dataset_STT/audio/uz/train/unpacked"
        # Yoki to'liq yo'li: tar_path = "Dataset_STT/audio/uz/train/train.tar"
        # "unpacked" papkani o'zingiz oldindan tar -xvf bilan yaratishingiz kerak.
        # Yoki tarfile moduli bilan python ichida extraction qilishingiz mumkin.
        
        # Soddalik uchun, tar ni allaqachon manual ravishda unpack qildik deb olamiz:
        audio_folder = tar_path.replace(".tar", "_unpacked")  
        # misol: "Dataset_STT/audio/uz/train/train_unpacked"
        
        # Keyin 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):
                if limit is not None and idx >= limit:
                    break  # 10k dan oshsa, to'xtaymiz
                
                audio_id = row["id"]
                mp3_file = audio_id + ".mp3"
                mp3_path = os.path.join(audio_folder, mp3_file)
                
                yield idx, {
                    "id": audio_id,
                    "audio": mp3_path,
                    "sentence": row["sentence"],
                    "duration": float(row["duration"]),
                    "age": row["age"],
                    "gender": row["gender"],
                    "accents": row["accents"],
                    "locale": row["locale"],
                }