Voxtral TTS code training skips encoder

Clever hack, but the real story is narrower than the headline suggests: it demonstrates that you can back into Voxtral-style codes without training a fresh codec from scratch. That makes it interesting for experimentation, reverse engineering, and small-scale audio workflows, but not yet a drop-in path for production TTS.

• –Uses differentiable optimization over discrete bottleneck codes, so the problem becomes “fit the codes” rather than “train a new encoder”
• –The repo claims workable results on CPU/MPS with about an hour of training on a Mac, which lowers the barrier for testing
• –Scope looks limited to reconstructing a known audio sample, so this is more proof-of-concept than a robust codec
• –If the approach holds up more broadly, it could be useful for codec inversion, audio token research, and Voxtral internals exploration