alvanlii/cantonese-youtube-transcription-chunked

---
dataset_info:
features:
- name: audio
dtype:
audio:
sampling_rate: 16000
- name: labels_1
dtype: string
- name: labels_2
dtype: string
- name: channel
dtype: string
- name: title
dtype: string
splits:
- name: train
num_bytes: 189408828505.48
num_examples: 516022
download_size: 185225339025
dataset_size: 189408828505.48
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

## Cantonese Youtube Pseudo-Transcription Dataset
- This dataset contains 2000+ hours of Cantonese audio from YouTube.
- Audio is transcribed using `simonl0909/whisper-large-v2-cantonese` (labels_1) and `Scrya/whisper-large-v2-cantonese` (labels_2) with speculative-decoding (`alvanlii/whisper-small-cantonese`).
- All audio files are truncated to a maximum of 30 seconds

#### TO-DOs
- [ ] Split audio based on speakers
- [ ] More data

This dataset contains over 2000 hours of Cantonese audio from YouTube, transcribed using `simonl0909/whisper-large-v2-cantonese` and `Scrya/whisper-large-v2-cantonese` models with speculative-decoding. All audio files are truncated to a maximum of 30 seconds. The dataset features include audio, labels, channel, and title, and provides a training set.

原始信息汇总

粤语YouTube伪转录数据集

数据集信息

• 特征:
  • audio: 音频文件，采样率为16000Hz。
  • labels_1: 字符串类型，使用simonl0909/whisper-large-v2-cantonese进行转录。
  • labels_2: 字符串类型，使用Scrya/whisper-large-v2-cantonese进行转录。
  • channel: 字符串类型。
  • title: 字符串类型。
• 分割:
  • train: 训练集，包含516022个样本，总大小为189408828505.48字节。
• 下载大小: 185225339025字节。
• 数据集大小: 189408828505.48字节。
• 配置:
  • default: 数据文件路径为data/train-*。

数据集描述

• 该数据集包含超过2000小时的粤语YouTube音频。
• 音频使用simonl0909/whisper-large-v2-cantonese和Scrya/whisper-large-v2-cantonese进行转录，并使用alvanlii/whisper-small-cantonese进行推测解码。
• 所有音频文件被截断至最长30秒。

待办事项

• [ ] 根据说话者分割音频。
• [ ] 增加更多数据。