Awesome
This is a reimplementaion of the neural vocoder in DIFFWAVE: A VERSATILE DIFFUSION MODEL FOR AUDIO SYNTHESIS.
Usage:
-
To continue training the model, run
python distributed_train.py -c config_${channel}.json
, where${channel}
can be either64
or128
. -
To retrain the model, change the parameter
ckpt_iter
in the correspondingjson
file to-1
and use the above command. -
To generate audio, run
python inference.py -c config_${channel}.json -cond ${conditioner_name}
. For example, if the name of the mel spectrogram isLJ001-0001.wav.pt
, then${conditioner_name}
isLJ001-0001
. Provided mel spectrograms includeLJ001-0001
throughLJ001-0186
. -
Note, you may need to carefully adjust some parameters in the
json
file, such asdata_path
andbatch_size_per_gpu
.