KittenTTS 2 review: voice cloning vs ElevenLabs

KittenTTS 2 adds voice cloning to a substantially larger speech model. The presenter tests it locally with a 10-second recording, while the project claims it can clone a voice from five seconds. The demonstrated setup calls for Python 3.10 or newer. Whisper transcribes the reference clip, and a speech language model generates audio tokens for the S3GEN vocoder from Chatterbox Turbo. Emotion tags and laughter remain beta features.

For the reviewed build, the presenter describes a 1.7-billion-parameter model with a default download of about 950 MB, plus smaller and full variants around 470 MB and 3.5 GB. On the presenter's Mac, the Python implementation falls back to CPU rather than using the Apple GPU. He cites documentation reporting generation two to three times slower than real time. A separate C++ port claims real-time CPU output, but he finds no published performance figures to support it.

The comparison with ElevenLabs centers on keeping voice recordings on the user's machine. Licensing complicates commercial deployment: the code and pip package retain Apache 2.0, but the weights use the revocable Stellan Labs Community License. The presenter describes registration, attribution and revenue or funding limits on free commercial use.

The presenter reports fragile non-English voices and conflicting language counts in the documentation he reviewed. That build requires normalization off for non-English generation. Automatic chunking addresses long passages that can repeat or cut off.