RealtimeSTT
Low-latency speech-to-text library with voice activity detection and wake words.
Build voice assistants, dictation tools, and streaming servers with minimal code.
RealtimeSTT is a Python library providing voice activity detection, fast transcription, and wake word activation. It suits developers building interactive voice applications or custom streaming servers.
What it is
The library captures audio streams, detects speech using WebRTC or Silero VAD, and transcribes audio via engines like faster-whisper or sherpa-onnx. It supports optional wake word detection through Porcupine or OpenWakeWord and includes a packaged production FastAPI server.
Who it's for
It suits Python developers building local voice interfaces, dictation software, or browser streaming backends. It does not suit teams seeking non-Python stacks or zero-dependency setups, as it requires audio drivers and heavy machine learning models.
Key features
Voice Activity Detection
Detects speech automatically using WebRTC VAD and Silero VAD.
Wake Word Activation
Supports optional wake word triggering via Porcupine or OpenWakeWord.
Multiple Engine Support
Runs on faster-whisper, sherpa-onnx, Moonshine, and other selectable backends.
Production FastAPI Server
Includes an optional remote server with versioned HTTP and WebSocket endpoints.
Up and running in 3 steps
Our honest verdict
What's great
- Provides direct microphone input or application-fed external audio chunks.
- Includes a packaged production FastAPI server with session isolation and authentication.
- Supports CPU-optimized streaming pipelines using sherpa-onnx models.
- Offers event callbacks for recording, VAD, realtime text, and wake word states.
Where it falls short
- Requires platform-specific system dependencies like PortAudio headers on Linux and macOS.
- Heavy machine learning models demand significant local compute resources or GPU hardware.
- Python 3.13 and newer are not yet supported release targets.
Pricing & license
Free under its MIT licence. There are no paid hosted options mentioned in the repository, so the only cost is running the hardware and models yourself.
Open remote AI-training roles
876 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
How do I feed audio from an external file or stream instead of a microphone?
Initialize the recorder with `use_microphone=False` and feed 16-bit mono PCM chunks using `feed_audio()`.
Can I run RealtimeSTT on a CPU-only production server?
Yes, the README recommends pairing sherpa-onnx-nemotron for live hypotheses with sherpa-onnx-nemo-parakeet for final transcripts on CPU.
