Pocket-TTS Server Gets Stuck in Cloning Mode

The issue reads less like a model bug and more like a product UX trap: catalog voices and cloning inputs are being conflated, so a user can accidentally enter the wrong inference path.

• –Kyutai’s official Pocket-TTS docs position voice cloning as optional, while the catalog voices are meant to work without downloading cloning weights.
• –Uploading `casual.wav` looks like a cloning prompt, not a voice-selection action, which is why the app tries to hydrate voice-state assets.
• –The practical fix is to choose a built-in voice explicitly in the voice selector or by `voice_url`, rather than uploading a reference WAV.
• –This is a good reminder that local AI tools need clearer mode separation; when the UI hides that boundary, users will assume a missing dependency instead of a configuration mistake.
• –For developers, the real risk is support churn: a single ambiguous control can make a working open-source model feel broken.