facodec

Low-bitrate factorized speech codec on top of FACodec: FSQ distillation to replace RVQ (WER, mel and speaker similarity vs bitrate, with bootstrap CIs) and a learned reverberation/noise "environment" factor (partial result, with limitations).

Autoregressive speech prediction with EnCodec and FACodec token language models: 2s context -> 1s future speech, with STOI/PESQ/DNSMOS analysis of where those metrics disagree.