github kyutai-labs/pocket-tts v3.2.0

5 hours ago

The french language has been updated, we now have "french" and a better "french_24l"!

What's Changed

  • Always end the prompt with sentence-final punctuation by @manu-lm in #296
  • README: add Pocket-TTS-LiteRT (Android GPU) to alternative implementations by @john-rocky in #295
  • Add pocket-tts-farsi to community-trained models by @mallahyari in #298
  • Add Indonesian 24L community model to README by @anak10thn in #294
  • Decode every queued latent per call in the decoder thread by @manu-lm in #300
  • Add pocket-tts-estonian to community-trained models by @cbentes in #301
  • README: collapse community model commands into
    Details blocks by @manu-lm in #302
  • Add dependency cooldown (exclude-newer = 7 days) by @gabrieldemarmiesse in #303
  • Bump the Indonesian model to its retrained weights by @anak10thn in #306
  • Add capitalize_first_letter so phoneme models are not capitalized by @mallahyari in #307
  • Length-bucketed batches by default by @manu-lm in #309
  • Add Pocket TTS Cymraeg (Welsh) community model by @dwain-barnes in #310
  • english_2026-09: robustness fine-tune (mid-word prompt cuts, unpunctuated text) as the default English model by @manu-lm in #311
  • Add Pocket TTS Polish (6 layers) community model by @shefowl in #315
  • Update README.md by @manukyutai in #316
  • Training: Muon as an optional optimizer by @manu-lm in #314
  • Stop generation when nobody is listening anymore by @bfulton in #228
  • Load tokenizers from tokenizer.json, keeping .model support by @manu-lm in #317
  • Refuse manifests without word alignments at startup by @manu-lm in #318
  • Ignore EOS on the first 6 generated frames by @manu-lm in #319
  • Keep the data loaders alive through SIGTERM so the trainer can checkpoint by @manu-lm in #320
  • French: 6-layer model and a new 24-layer model by @manu-lm in #321

New Contributors

Full Changelog: v3.1.0...v3.2.0

Don't miss a new pocket-tts release

NewReleases is sending notifications on new releases.