metadata
dataset_info:
features:
- name: audio
dtype: audio
- name: transcript
dtype: string
splits:
- name: train
num_bytes: 417515940220.32
num_examples: 4330930
download_size: 700809728224
dataset_size: 417515940220.32
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- text-to-speech
- automatic-speech-recognition
size_categories:
- 1M<n<10M
Cantonese Audio TTS Dataset
This dataset contains alvanlii/cantonese-radio, alvanlii/cantonese-youtube, plus a dataset of equal size. It is catered towards TTS (text-to-speech) use cases, more than the 2 previously published datasets, as there is more extensive filtering and audio enhancement. For speaker labelling, you can use speaker embedding models like Nvidia's TitaNet
Filtered out:
- Overlapped voices, detected using pyannote/speaker-diarization-3.1
- Music, detected using a custom model
- Languages not Chinese or English
- Transcriptions that are mostly English
- Deduplication using MinHashLSH
- Empty audio, using VAD models
Audios are
- enhanced/denoised via sarulab-speech/sidon-v0.1 and ClearVoice
- transcribed using alvanlii/whisper-small-cantonese and stepfun-ai/Step-Audio-2-mini