aask1357 / hilcodec

High fidelity, lightweight, end-to-end, streaming, convolution-based neural audio codec
MIT License
74 stars 7 forks source link

checkpoints would be public? #2

Closed deyituo closed 5 months ago

deyituo commented 5 months ago

I find onnx model with 8 qualitizers is public, but 12 in the configs, would you public the full checkpoints?

aask1357 commented 5 months ago

We added the hilcodec_music checkpoint (a model trained on general audio), which has 12 quantizers. hilcodec_speech (a model trained on clean speech) has 8 quantizers, which is the default configuration. This is because we found that for clean speech only, 8 quantizers (6kbps @ 24kHz) is enough. We'll update the configs/hilcodec_speech.yaml file to have 8 quantizers as default. Thank you!

deyituo commented 5 months ago

Thanks for your reply!