metadata
license: mit
language:
- ja
base_model:
- Akjava/matcha-tts_ja_100speakers_group003f-CL-V1
datasets:
- Akjava/ja004_speech_common-voice_22khz
- Akjava/ja005_speech_common-voice_22khz
- Akjava/ja006_speech_common-voice_22khz
- Akjava/ja007_speech_common-voice_22khz
- Akjava/ja008_speech_common-voice_22khz
- Akjava/ja009_speech_common-voice_22khz
This model has 100 slots and 6 speaker use them
ITA-Rectaion-010:家具商人のフィシェルは、荷車と仔馬を貸してくれた。(kagushooniNnofisheruwa,nigurumatokoumaokashItekureta.)
spk00:ja004
spk01:ja005
spk02:ja006
spk03:ja007
spk04:ja008
spk05:ja009
Training
some how machine would freeze with 80 batch size ,change to 40
Stage1(2024-09-12_12-07-55)
maybe it need more training time,all voices are robotic
Stage2(2024-09-12_22-09-01)
most of them improved
spk0 - differenct intonation than v2
spk1 - v2 is better
spk2 - good
spk3 - good
spk4 - good
spk5 - Not so good .miss training dataset
Importan checkpoint
5709 - latest