RealtimeSTT

RealtimeSTT wires the pieces of a live transcription pipeline into one Python library: voice activity detection through WebRTC VAD and Silero VAD, optional wake word activation through Porcupine or OpenWakeWord, and streaming transcription with faster-whisper as the recommended engine, with OpenAI Whisper, Moonshine, sherpa-onnx, Kroko-ONNX, Parakeet, and FunASR available as alternatives. The recorder listens on a microphone or accepts external audio streams and emits partial text in real time, which is why the library is commonly used as the ears of LLM voice agents, often paired with the same author's RealtimeTTS for the reply path. Installation is pip install RealtimeSTT with the faster-whisper extra on Python 3.11 or newer, with PortAudio headers needed on Linux; CPU-only operation works with small models, and a CUDA GPU is recommended for realtime accuracy with larger ones. A FastAPI browser server example demonstrates multi-user streaming. The project is MIT licensed, has around 10,000 GitHub stars, and everything, including wake words and VAD, runs locally.