In the paper, it mentions "using a 32-layer residual vector quantizer (RVQ)". Does this refer to MossAudioTokenizerResidualVQ (RVQ) or MossAudioTokenizerResidualLFQ (RLFQ) in the code? How do these two quantizers differ in audio reconstruction quality and downstream task performance (e.g., ASR, classification)?
In the paper, it mentions "using a 32-layer residual vector quantizer (RVQ)". Does this refer to MossAudioTokenizerResidualVQ (RVQ) or MossAudioTokenizerResidualLFQ (RLFQ) in the code? How do these two quantizers differ in audio reconstruction quality and downstream task performance (e.g., ASR, classification)?