Thank you for the open-weights release — the quality and the architecture write-up are both impressive.
The release currently ships only the decode-side components: the Global/Local LLMs, RVQ depth decoder, flow-matching module, Flow-VAE/vocoder (and dav.pth, which does include its codec encoder). However, the 8-layer RVQ music tokenizer's encoder — the component that maps a waveform to the semantic (layer-1) and acoustic (layers 2–8) token sequences — does not appear to be included.
Without it, fine-tuning the Global LLM is not possible, because there is no way to produce target semantic-token sequences from one's own audio data. Our use case is research on melody/harmony-conditioned generation: fine-tuning the Global LLM on paired symbolic+audio data while keeping the entire synthesis stack frozen.
Questions:
- Are there plans to release the tokenizer encoder weights (and preprocessing code), similar to how other open music/speech releases have shipped their codecs?
- If not, could you share the tokenizer's training configuration so the community can attempt a faithful reimplementation?
Thanks again for pushing open music generation forward.
Thank you for the open-weights release — the quality and the architecture write-up are both impressive.
The release currently ships only the decode-side components: the Global/Local LLMs, RVQ depth decoder, flow-matching module, Flow-VAE/vocoder (and
dav.pth, which does include its codec encoder). However, the 8-layer RVQ music tokenizer's encoder — the component that maps a waveform to the semantic (layer-1) and acoustic (layers 2–8) token sequences — does not appear to be included.Without it, fine-tuning the Global LLM is not possible, because there is no way to produce target semantic-token sequences from one's own audio data. Our use case is research on melody/harmony-conditioned generation: fine-tuning the Global LLM on paired symbolic+audio data while keeping the entire synthesis stack frozen.
Questions:
Thanks again for pushing open music generation forward.