RealtimeTTS

RealtimeTTS solves a narrow but common problem: turning strings, generators, and LLM token streams into audible speech with as little latency as possible. The Python library sits as a glue layer above more than 20 engines, from free system voices, gTTS, and Edge through cloud APIs like OpenAI, Azure, ElevenLabs, Cartesia, and MiniMax to local neural models including Coqui XTTS, Piper, Kokoro, StyleTTS, Parler, Chatterbox, and Orpheus, with configurable fallback chains that switch engines automatically when one fails. Text can keep streaming in while audio is already playing, and the API covers pause, resume, stop, WAV file output, audio callbacks, voice switching, multiple output devices, and asynchronous playback. Installation is pip install with per-engine extras, such as realtimetts[system], realtimetts[coqui], or realtimetts[openai]. Hardware needs depend entirely on the chosen engine; system voices run anywhere while local neural engines benefit from CUDA acceleration. The MIT license permits commercial use, and with about 4,000 GitHub stars the library has become a standard way to pipe streaming LLM output into speech.