Chatterbox

Chatterbox is Resemble AI's open-source text-to-speech model family, MIT-licensed. The lineup covers three variants: the 500M standard model with classifier-free guidance and exaggeration tuning for creative control, Chatterbox Multilingual (500M, 23 languages), and Chatterbox Turbo (350M, distilled one-step decoder for ultra-low-latency real-time synthesis). It was the first open TTS model with emotion exaggeration control, ships with the PerTh implicit watermarker for responsible deployment, and in Podonos blind evaluations has been preferred over ElevenLabs and Cartesia voices.

The honest picture: as with all open TTS, quality depends on your prompt and reference audio — the headline eval wins are under controlled conditions, not a guarantee on every voice. The Turbo variant trades some expressiveness for speed. And the watermarking, while responsible, means outputs are detectable as synthetic, which matters if you need "clean" audio. Resemble AI also sells a hosted API, so the open models are the community tier of a commercial funnel.