Inquiry on Semantic Richness and Acoustic Fidelity Variation with n_q in XCodec, and Challenges in Scaling to 44kHz

[LiuZH-19]

Great work!
I would like to inquire if there are any results available regarding the variation of semantic richness and acoustic fidelity as the number of n_q changes in XCodec. Specifically, I am interested in understanding how these two factors (semantic richness and acoustic fidelity) behave as n_q is increased or decreased.

Additionally, I have observed that XCodec operates at a sampling rate of 16kHz, and the reconstructed WAV files lose many acoustic details compared to raw 44.1kHz audio. I am curious about the challenges involved when applying XCodec's technology to a 44kHz codec. For instance, would it be feasible to enhance the DAC by integrating Hubert-based representations?

Any insights or experiences would be greatly appreciated.

Thank you!

[humblenginr]

Hey man,
Saw your work on songgen, impressed by the results but I couldn't get it to produce same results using your last checkpoint

Any tips or do you think more compute will get us to the level that you showed in your samples?

A bit confused if its actually possible to train a suno level model using the architecture you described