A local Voice API for your cloned voices, trained from your own samples

Voice API: your cloned voices, served to other programs on this device — a new Voice API space (⌘7, a Pro feature) runs an ElevenLabs-compatible API on your computer, so code written for the hosted service works here by changing only the base URL. Speech is made on this device by the same engines as Voice Clone. Start server in the header turns it on and off; it starts off. The space has three tabs: Voices, Playground and API access. Scripts can speak text, stream speech sentence by sentence, and list, add, edit and delete voices. Any eleven_* model id speaks with the voice's own engine. MP3 and Opus requests come back as WAV at the rate you asked for, because no MP3 encoder ships with the app.

Train a voice from up to five samples — Train a voice on the Voices tab takes Record new, Saved dictations or Import files. Record new gives you a passage to read aloud and records it without the echo cancellation and noise suppression a clone would otherwise copy. The listening overlay shows while you read. The sheet offers up to three of the clearest excerpts, best first. Listen, and keep the one that sounds most like you. Qwen voices need the excerpt's exact words: Transcribe excerpt writes them with local Whisper, and you can correct any word before Save voice. A saved voice keeps its own copy of the excerpt, so deleting a dictation in Review does not break a script that calls the voice.

See how close each clone sounds — with the speaker model installed (26 MB, the one Listen uses), the Voices tab scores every voice against its reference as very close, close, recognisable or not this voice, with suggestions when the score is low. New voices are tuned in the background for the closest result, with a clone-check take to play beside the reference. Tune voice runs it again; Chatterbox voices are tuned only when you press it.

Playground for trying requests — send a request through the same path a script uses, even with the server off, and see the status, headers, similarity and audio a client would get, with matching cURL, Python and JavaScript snippets. Compare engines speaks the same text with every engine installed for the voice and scores each take. Save WAV to Downloads keeps a result.

A newer Pocket TTS model hears your whole reference — the one-time download is now 146 MB, and the model listens to the entire excerpt rather than only its first second, so clones sound closer. If you already have the earlier model it keeps working, and the Voices tab offers the newer one.

Updates follow your license's update window — a Pro license whose year of updates has ended is asked to renew before a new version installs, and a renewal counts on the next check rather than the next launch. Free installs update as before.