Plans for 24kHz output? #3
Description
Fantastic work. I’ve been testing MingTok-Audio and it performs exceptionally well in reconstruction quality and naturalness in my pipelines. The design choices (continuous latents) make it very practical for multiple setups.
However, one limitation that stands out is the 16 kHz constraint. Several recent SOTA discrete tokenizers for speech are trained at 24 kHz (and many toolchains assume 24 kHz for high-quality speech and 48 kHz for broader audio), which improves high-frequency cues (sibilants, air/brightness), prosodic detail, and downstream fidelity.
Is a 24+ kHz variant planned? If so, any rough timeline or milestones you can share? If not, this is another reason to release training code to the community.