PocketTTS (Kyutai) as a fast CPU-only engine?

[paoloantinori]

Context

A friend introduced me to PocketTTS — Kyutai Labs' 100M-parameter TTS with voice cloning that runs real-time on CPU, no GPU needed. I found myself deploying it alongside OmniVoice for latency-sensitive tasks (MCP agent responses, quick previews) where OmniVoice's quality is overkill and the GPU warm-up + inference time is the bottleneck.

It's MIT-licensed, pip-installable, with a simple API:

from pocket_tts import TTSModel
tts = TTSModel.load_model()
voice = tts.get_state_for_audio_prompt("./ref.wav")
audio = tts.generate_audio(voice, "Hello world")

~6 languages (EN + 5 from April), voice cloning from a 20s clip.

The question

No pressure to adopt it — they serve different tiers (646 languages + dubbing + dictation + design vs. fast lightweight CPU TTS). But would a PocketTTS engine option be interesting in OmniVoice? Two use cases:

1. Fast draft / preview — instant TTS on CPU while the full-quality render runs on GPU.
2. GPU-less deployments — users without CUDA/MPS could still run OmniVoice.

Happy to spike the integration if it's interesting. If not, no worries — I'll keep running them side by side.