AEmotionStudio/pocket-tts-models

Pocket TTS — MAESTRO mirror

Verbatim mirror of the six per-language Pocket TTS checkpoints MAESTRO ships, taken from Kyutai's ungated release kyutai/pocket-tts-without-voice-cloning (every file byte-identical, sha256-verified — no conversion). This repo exists so MAESTRO's in-app downloader has a stable, self-contained source per language.

Attribution (CC-BY-4.0)

The model weights are © Kyutai and licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Credit: Pocket TTS by Kyutai — Manu Orsini, Simon Rouard, Gabriel de Marmiesse, Václav Volhejn, Neil Zeghidour, Alexandre Défossez. Paper: https://arxiv.org/abs/2509.06926 · Tech report: https://kyutai.org/blog/2026-01-13-pocket-tts No changes were made to the files.

Layout

languages/
  english/         6-layer 100M model + sentencepiece tokenizer + 26 voice embeddings
  french_24l/      24-layer model (upstream ships no smaller French checkpoint)
  german/  spanish/  italian/  portuguese/

What this mirror deliberately is NOT

Kyutai distributes the voice-cloning-capable checkpoints behind a gate with a prohibited-use agreement (voice impersonation without consent, deception, etc.). This mirror follows their split: the mimi.encoder* tensors here are zeroed (exactly Kyutai's own ungated release), so these files support the 26 preset voices but cannot clone. MAESTRO users who want cloning fetch the intact checkpoint directly from the gated kyutai/pocket-tts repo with their own HF token after accepting Kyutai's terms.

Voice embedding provenance

The 26 preset voice embeddings derive from clips catalogued (with per-voice licenses) in kyutai/tts-voices: mostly VCTK (CC BY 4.0), Unmute voice donations / Voice-Zero / Common Voice (CC0), and Kyutai recordings. Two voices derive from non-commercial source datasets — cosette (Expresso, CC BY-NC 4.0) and jean (EARS, CC BY-NC 4.0) — MAESTRO labels these in-app and recommends the other 24 for commercial work.

Prohibited use (from upstream)

Use must comply with all applicable laws and must not involve voice impersonation or cloning without explicit and lawful consent, misinformation or deception, or the generation of unlawful, harmful, libelous, abusive, discriminatory, or privacy-invasive content.