ChatTTS

Sweet spot: a developer or researcher who wants a locally-run TTS with conversational realism and is comfortable with a research-only license. On a mid-range GPU it generates audio quickly enough for interactive use, and the <laugh> / <break> tag system gives you real creative control. Failure modes. The licensing question is the single biggest blocker for shipping anything commercial — do not assume MIT. Voice quality, while strong for open-source, still trails top commercial TTS (ElevenLabs, Play.ht) on pure naturalness and consistency across long passages. Support is community-driven; expect rough edges. What to pilot. Run a 30-second demo locally. If the voice quality clears your bar and the license fits your use case, plan on wrapping it with a lightweight FastAPI layer and caching audio for repeat phrases. If you need commercial-grade output at scale, budget for a paid TTS — the cost of licensing negotiation exceeds what a commercial API charges.

ChatTTS is an open-source TTS model from 2noise, specifically trained on conversational data rather than the audiobook-style corpora that dominate most open TTS systems. The result is a voice that includes natural disfluencies, laughter, stress patterns, and turn-taking cues — useful for podcast generation, game NPCs, and AI agents that need to sound less like a radio announcer. It supports Chinese and English, emits prosody tags inline (<laugh>, <break>, <stress>), and runs reasonably fast on a single GPU. The model weights are open under a research license, and a demo is available at 2noise.com. A growing set of community wrappers adds streaming, voice cloning, and REST API shims on top of the base model. ChatTTS is not a commercial product — there is no hosted SaaS or official API — but it is one of the most downloaded open TTS checkpoints on HuggingFace and a foundational component in several open-source voice-agent projects.