|
|
--- |
|
|
license: mit |
|
|
language: |
|
|
- ja |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
--- |
|
|
## TTS Training Detail |
|
|
[Size-Category](https://github.com/akjava/Matcha-TTS-Japanese/wiki/Size-Category-%E3%82%B5%E3%82%A4%E3%82%BA%E3%82%AB%E3%83%86%E3%82%B4%E3%83%AA%E3%83%BC) - Large(500-5000):1230 files |
|
|
|
|
|
|
|
|
|
|
|
## Audio Detail |
|
|
- 22050hz mono wav - for modern tts (original are 32khz mono mp3) |
|
|
- vadded 200ms |
|
|
- Volume-level:OK - no need to gain. |
|
|
- Noise-level:OK - no need to denoise |
|
|
## Creating Detail |
|
|
- Based [Common Voice](https://huggingface.co/papers/1912.06670)-[16_1 ja](https://huggingface.co/datasets/mozilla-foundation/common_voice_16_1) |
|
|
- Audios are vadded with [silero-vad](https://github.com/snakers4/silero-vad) |
|
|
- Text are cleaned simply by openjtalk.(this mean possible contain miss reading and miss converting) |
|
|
|
|
|
## Don't attempt to identify the identity of the speaker in the voices. |
|
|
I agree to download original **public domain** dataset. |
|
|
|
