README.md

metadata

dataset_info:
  features:
    - name: filename
      dtype:
        audio:
          sampling_rate: 16000
    - name: url
      dtype: string
  splits:
    - name: train
      num_bytes: 1068464089483.938
      num_examples: 59879
  download_size: 16395869337
  dataset_size: 1068464089483.938

Crawl Youtube

Malaysian and Singaporean youtube channels, total up to 60k audio files with total 18.7k hours.

URLs data at https://github.com/mesolitica/malaya-speech/tree/master/data/youtube/data

Notebooks at https://github.com/mesolitica/malaya-speech/tree/master/data/youtube

How to load the data efficiently?

import pandas as pd
import json
from datasets import Audio
from torch.utils.data import DataLoader, Dataset

chunks = 30
sr = 16000

class Train(Dataset):
    def __init__(self, indices, maxlen_cache_df=5, maxlen_cache_audio=50):
        self.indices = {}
        for k, v in indices.items():
            for i in range(int(k), v['start'] + v['end'], 1):
                self.indices[i] = v

        self.max_index = len(self.indices)
        self.cache_df = {}
        self.cache_audio = {}
        self.maxlen_cache_df = maxlen_cache_df
        self.maxlen_cache_audio = maxlen_cache_audio
        self.audio = Audio(sampling_rate=16000)

    def __len__(self):
        return self.max_index

    def __getitem__(self, item):
        if item < 0:
            item = self.max_index + item

        v = self.indices[item]

        key_row = f"{v['filename']}-{v['i']}"
        chunk_index = item - v['start']

        if key_row not in self.cache_audio:

            if v['filename'] not in self.cache_df:
                df = pd.read_parquet(v['filename'])
                if len(self.cache_df) >= self.maxlen_cache_df:
                    keys = list(self.cache_df.keys())
                    self.cache_df.pop(sorted(keys)[0], None)
                self.cache_df[v['filename']] = df
            else:
                df = self.cache_df[v['filename']]

            row = df.iloc[int(v['i'])]
            audio = self.audio.decode_example(self.audio.encode_example(row['filename']))
            if len(self.cache_audio) >= self.maxlen_cache_audio:
                keys = list(self.cache_audio.keys())
                self.cache_audio.pop(sorted(keys)[0], None)
            self.cache_audio[key_row] = audio

        else:
            audio = self.cache_audio[key_row]

        return {
            'array': audio['array'][(chunks * sr) * chunk_index: (chunks * sr) * (chunk_index + 1)]
        }

with open('crawl-youtube-global-indices.json') as fopen:
    global_indices = json.load(fopen)

train = Train(global_indices)
train[0]

{'array': array([ 0.        ,  0.        ,  0.        , ..., -0.00845753,
         0.00168016, -0.00606468])}

This is global hashing indices if the audio chunked with 30 seconds, read more at https://github.com/mesolitica/malaysian-dataset/tree/master/speech-to-text-semisupervised/pseudolabel-whisper

Licensing

All the videos, songs, images, and graphics used in the video belong to their respective owners and I does not claim any right over them.

Copyright Disclaimer under section 107 of the Copyright Act of 1976, allowance is made for "fair use" for purposes such as criticism, comment, news reporting, teaching, scholarship, education and research. Fair use is a use permitted by copyright statute that might otherwise be infringing.