ml-superb-subset.py

import csv
import glob
import os
import textwrap
from dataclasses import dataclass

import tqdm

import datasets
from datasets.tasks import AutomaticSpeechRecognition

from typing import List

LANGUAGES = ["afr", "amh", "azz", "nbl", "nso", "sot", "ssw", "swa", "tos", "tsn", "tso", "ven", "wol", "xho", "xty", "zul"]

class MLSuperbConfig(datasets.BuilderConfig):
    """BuilderConfig for Superb."""

    def __init__(self, name, **kwargs):
        super(MLSuperbConfig, self).__init__(name=name, version=datasets.Version("2.19.0"), **kwargs)




class MLSuperb(datasets.GeneratorBasedBuilder):
    DEFAULT_WRITER_BATCH_SIZE = 1000
    URL = "https://224sh3.s3.amazonaws.com/ml_superb_subset.zip"

    # BUILDER_CONFIG_CLASS = MLSuperbConfig

    BUILDER_CONFIGS = [
        MLSuperbConfig(
            name=lang,
        )
        for lang in LANGUAGES
    ]

    def _info(self):
        features = datasets.Features(
            {
                "audio": datasets.Value("string"),
                "sentence": datasets.Value("string"),
            }
        )

        return datasets.DatasetInfo(
            features=features,
            supervised_keys=None,
            version=self.config.version,
        )



    def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
        urls_to_download = self.URL
        downloaded_files = dl_manager.download_and_extract(urls_to_download)
        downloaded_files = downloaded_files + "/ml_superb_subset/" + self.config.name
        # downloaded_files = "./ml_superb_subset/" + self.config.name
        # downloaded_files = "ml-superb-subset" "/ml_superb_subset/" + self.config.name


        splits = ("train10min", "train1hr", "dev", "test")
        split_to_filename = {
            "train10min": 'transcript_10min_train.txt',
            "train1hr": 'transcript_1h_train.txt',
            "dev": 'transcript_10min_dev.txt',
            "test": 'transcript_10min_test.txt',
        }
        split_generators = []
        split_names = {
            "train10min": datasets.Split.TRAIN,
            "train1hr": datasets.Split.TRAIN,
            "dev": datasets.Split.VALIDATION,
            "test": datasets.Split.TEST,
        }
        for split in splits:
            split_generators.append(
                datasets.SplitGenerator(
                    name=split,
                    gen_kwargs={
                        'wavs_path' : downloaded_files + "/wav/",
                        "transcript_path": downloaded_files + "/" + split_to_filename[split],
                    },
                ),
            )

        return split_generators


    def _generate_examples(self, wavs_path, transcript_path):
        data_fields = list(self._info().features.keys())
        metadata = {}
        with open(transcript_path, encoding="utf-8") as f:
            reader = csv.reader(f, delimiter="\t", quoting=csv.QUOTE_NONE)
            if len(next(reader)) == 1:
                reader = csv.reader(f, delimiter=" ", quoting=csv.QUOTE_NONE)
                for row in reader:
                    id_ = row[0]
                    if not row[0].endswith(".wav"):
                        row[0] += ".wav"
                    metadata[row[0]] = " ".join(row[2:])

                    yield id_, {
                        "audio": wavs_path + row[0],
                        "sentence": " ".join(row[2:]),
                        "id": id_,
                    }
            else:
                for row in reader:
                    # print(row)
                    id_ = row[0]
                    if not row[0].endswith(".wav"):
                        row[0] += ".wav"
                    metadata[row[0]] = row[-1]

                    yield id_, {
                            "audio": wavs_path + row[0],
                            "sentence": row[-1],
                            "id": id_,
                        }