Artoria2e5
commented
1 year ago David Rowe from codec2 has already tried that:
- https://www.rowetel.com/?p=5966 LPCNet can be mated to Codec2 features.
- https://www.rowetel.com/?p=6639. In this second experiment he feeds LPCNet data from an already quantized stream of Codec 2. This ML-based decoding approach gives better sound: "Even the lpcnet_1300 decoder sounds better than Codec 2 at 2400."
But then it's probably better to just keep what the net already has. https://www.rowetel.com/?p=6639 tries a 1700 kbps quantized mode, independently from the current 1600 thing (but definitely similar!) I bet the step forward is continuing to squeeze more bits out of the quantizer. If you check https://www.rowetel.com/?cat=7, you can see Rowe doing a good amount of work on all kinds of vector quantisers.
Can I combine Codec2 with LPCNet to do ML experiments at 700bps?