RealtimeTTS is an open-source Python library that converts text to speech in real time, with latency low enough for conversational AI. Published on GitHub, it has earned about 4,043 stars and is under active development in 2026. It is designed for the way modern voice assistants work: text arrives token by token from a language model, and RealtimeTTS starts speaking the first sentence while the rest is still being generated.
The library intelligently splits incoming text streams into speakable chunks, synthesizes them through a pluggable engine, and plays or records the audio with minimal delay. Engines range from fully local neural voices such as Coqui XTTS to system voices and cloud services, and the architecture makes it straightforward to add new ones. Features like sentence-level streaming, audio recording to file, and playback controls make it a practical building block for voice interfaces, accessibility tools, and talking avatars. Developers can also adjust playback behavior, and the library handles interruptions gracefully when a user starts talking over the assistant.
RealtimeTTS is free and open source under the MIT license. Local engines cost nothing to run; if you plug in a commercial cloud voice engine, that provider’s own usage pricing applies, but the library itself is free.
RealtimeTTS is for developers building voice assistants, conversational agents, accessibility readers, or any application where an AI needs to speak naturally and immediately rather than waiting for a full response to be synthesized.