Easy to scale and Good control over latency and caching
No data length limit is required, such as 1s
cache and multiple speech models are available out of the box
{"wav": "/data/BAC009S0764W0121.wav"}
{"wav": "/data/BAC009S0764W0122.wav"}train_data = 'train.jsonl'
model_dir = 'vocos/exp/2025/0.1/transformer/'
tensorboard_dir = ${model_dir}/runs/
mkdir -p $model_dir $tensorboard_dir
torchrun --standalone --nnodes=1 --nproc_per_node=8 vocos/main.py -- \
--config vocos/configs/default.py \
--config.train_data=${train_data} \
--config.model_dir=${model_dir} \
--config.tensorboard_dir=${tensorboard_dir} \
--config.max_train_steps 1000000TODO:
- training
- training works
- check training process
- generator
- disc
- transformer discriminators
- distill
- resume
- stereo for music
- cqt loss
- dev benchmark etc
- infer
- offline
- chunk by chunk or frame by frame
- onnx
- exmple for: cosyvoice2 and transformer-vocos