dataset.py

# coding=utf-8
# Copyright 2023
#
# Licensed under the Apache License, Version 2.0 (the "License");
# http://www.apache.org/licenses/LICENSE-2.0
#
# Ushbu fayl Common Voice uslubidagi dataset loading script bo'lib,
# audio/uz/<split>/<split>.tar va transcript/uz/<split>/<split>.tsv
# fayllarni yuklab, audio+transkriptsiyani birlashtiradi.

import os
import csv
import json

import datasets
from datasets.utils.py_utils import size_str


# ------------------ 1. Metadata va sozlamalar ------------------
_CITATION = """\

@misc{yourcitation,

  title  = {Your STT dataset title},

  author = {You or your org},

  year   = {2023},

  url    = {https://huggingface.co/datasets/Elyordev/new_dataset_stt_audio}

}

"""

_DESCRIPTION = """\

Bu dataset mp3 formatdagi audio fayllar va tsv metadata fayllardan iborat.

Papka tuzilishi Common Voice uslubiga o'xshash:

audio/uz/[train|validation|test]/*.tar va transcript/uz/[train|validation|test]/*.tsv

"""

_HOMEPAGE = "https://huggingface.co/datasets/Elyordev/new_dataset_stt_audio"
_LICENSE = "Apache License 2.0"

# Bitta til: "uz" (xohlasangiz ko'paytirishingiz mumkin)
LANGUAGES = {
    "uz": {
        "language_name": "Uzbek",
        "num_clips": None,      # Agar xohlasangiz, taxminiy klip sonini kiriting
        "num_speakers": None,
        "validated_hr": None,
        "total_hr": None,
        "size_bytes": None,
    },
}

# Bizda har bir splitda 1 dona tar shard bor deb faraz qilamiz
N_SHARDS = {
    "uz": {
        "train": 1,
        "validation": 1,
        "test": 1,
    }
}

# Asosiy URL: repodagi fayllarni resolve qilish uchun
_BASE_URL = "https://huggingface.co/datasets/Elyordev/new_dataset_stt_audio/resolve/main/"

# Audio fayl yo'li: audio/uz/<split>/<split>.tar
_AUDIO_URL = _BASE_URL + "audio/{lang}/{split}/{split}.tar"

# Transcript fayl yo'li: transcript/uz/<split>/<split>.tsv
_TRANSCRIPT_URL = _BASE_URL + "transcript/{lang}/{split}/{split}.tsv"


# ------------------ 2. Config klassi ------------------
class NewDatasetSTTAudioConfig(datasets.BuilderConfig):
    """Bitta config (masalan, 'uz') - xohlasangiz ko'proq tillarni ham qo'shishingiz mumkin."""
    def __init__(self, language, **kwargs):
        super().__init__(**kwargs)
        self.language = language
        self.num_clips = LANGUAGES[language]["num_clips"]
        self.num_speakers = LANGUAGES[language]["num_speakers"]
        self.validated_hr = LANGUAGES[language]["validated_hr"]
        self.total_hr = LANGUAGES[language]["total_hr"]
        self.size_bytes = LANGUAGES[language]["size_bytes"]
        self.size_human = size_str(self.size_bytes) if self.size_bytes else None


# ------------------ 3. Asosiy dataset builder ------------------
class NewDatasetSTTAudio(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        # Masalan, faqat "uz" config
        NewDatasetSTTAudioConfig(
            name="uz",
            version=datasets.Version("1.0.0"),
            description="Uzbek STT dataset with Common Voice-like structure",
            language="uz",
        ),
    ]
    DEFAULT_WRITER_BATCH_SIZE = 1000

    def _info(self):
        lang = self.config.language
        # O'zingiz xohlagancha izoh tuzishingiz mumkin
        description = (
            f"Common Voice uslubidagi dataset: til = {lang}. "
            f"{_DESCRIPTION}"
        )
        features = datasets.Features(
            {
                "id": datasets.Value("string"),
                "path": datasets.Value("string"),
                "audio": datasets.features.Audio(sampling_rate=16000),  # agar 16kHz bo'lsa
                "sentence": datasets.Value("string"),
                "age": datasets.Value("string"),
                "gender": datasets.Value("string"),
                "accents": datasets.Value("string"),
                "locale": datasets.Value("string"),
                "duration": datasets.Value("float"),  # agar tsv da float bo'lsa
            }
        )
        return datasets.DatasetInfo(
            description=description,
            features=features,
            supervised_keys=None,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
            version=self.config.version,
        )

    def _split_generators(self, dl_manager):
        """

        Common Voice misolida bo'lgani kabi:

        train, validation, test splitlari uchun tar va tsv fayllarni yuklaymiz.

        """
        lang = self.config.language
        n_shards = N_SHARDS[lang]  # {'train':1, 'validation':1, 'test':1}
        split_generators = []

        # Bizda splits = ["train", "validation", "test"]
        for split in ["train", "validation", "test"]:
            # Audio (tar) URL lar ro'yxati (har bir splitda bitta shard)
            audio_urls = [
                _AUDIO_URL.format(lang=lang, split=split, shard_idx=i)
                for i in range(n_shards[split])
            ]
            # .tar fayllarni yuklab olamiz
            audio_paths = dl_manager.download(audio_urls)

            # .tar fayllarni streaming yoki to'liq extract qilamiz
            # Common Voice 'iter_archive' orqali stream qiladi, lekin biz local_extracted qilsak ham bo'ladi
            local_extracted_archive_paths = []
            if not dl_manager.is_streaming:
                local_extracted_archive_paths = dl_manager.extract(audio_paths)

            # Transcript (tsv) URL
            transcript_url = _TRANSCRIPT_URL.format(lang=lang, split=split)
            transcript_path = dl_manager.download_and_extract(transcript_url)

            split_generators.append(
                datasets.SplitGenerator(
                    name=getattr(datasets.Split, split.upper()),
                    gen_kwargs={
                        "archives": [
                            dl_manager.iter_archive(path) for path in audio_paths
                        ],
                        "local_extracted_archive_paths": local_extracted_archive_paths,
                        "meta_path": transcript_path,
                    },
                )
            )

        return split_generators

    def _generate_examples(self, archives, local_extracted_archive_paths, meta_path):
        """

        Har bir split uchun:

        1) transcript .tsv faylni o'qish

        2) audio tar ichidagi fayllarni "archives" orqali iteratsiya qilish

        3) tsv'dagi 'path' bilan audio fayl nomini bog'lash

        4) natijada (key, example) qaytarish

        """
        # Tsv fayl (meta_path) ni o'qib, metadata lug'atini tuzamiz
        # formati: { "filename.mp3": { ... ustunlar ... } }
        metadata = {}
        with open(meta_path, encoding="utf-8") as f:
            reader = csv.DictReader(f, delimiter="\t")
            for row in reader:
                # Ehtiyot chorasi: agar path .mp3 bilan tugamasa, qo'shamiz
                if not row["path"].endswith(".mp3"):
                    row["path"] += ".mp3"
                metadata[row["path"]] = row

        # Endi tar fayllarni o'qish
        # Common Voice misolida bir splitda bir nechta shard bo'lishi mumkin, shuning uchun ro'yxat
        for shard_idx, archive in enumerate(archives):
            # archive = dl_manager.iter_archive(path) => (path_in_tar, fileobj) generator
            for path_in_tar, fileobj in archive:
                # Masalan, path_in_tar = "common_voice_uz_12345.mp3"
                _, filename = os.path.split(path_in_tar)
                if filename in metadata:
                    # Metadata qatorini olish
                    row = metadata[filename]
                    # local_extracted_archive_paths[shard_idx] => .tar fayl extract qilingan joy
                    # Agar to'liq extract bo'lmagan bo'lsa, to'g'ridan-to'g'ri bytes bilan ham ishlasa bo'ladi
                    # Common Voice rasmiy misolida 'result["audio"] = {"path": path, "bytes": file.read()}' qilingan

                    example = dict(row)
                    # "id" ustuni bo'lmasa, idx sifatida path_in_tar dan foydalansa bo'ladi
                    if "id" not in example:
                        example["id"] = filename

                    # Audio: tar fayl ichidan o'qilgan bytes
                    # Datasets "Audio" featuri "bytes" ni o'z-o'zidan tan olmaydi,
                    # lekin "path" + "bytes" berish uslubi Common Voice scriptida ishlatilgan
                    # Keyingi bosqichda decode qilinadi
                    example["audio"] = {
                        "path": path_in_tar,
                        "bytes": fileobj.read(),
                    }

                    # Qo'shimcha ustunlarni ham row dan olamiz:
                    # sentence, age, gender, accents, locale, duration, ...
                    # Agar bo'lmasa, bo'sh qiymat kiritiladi
                    # Biz "metadata"dan oldin dict(row) deb oldik, demak "example"da hamma ustun bor.

                    yield path_in_tar, example