RealtimeSTT is a Python speech-to-text library that provides voice activity detection, fast transcription, wake word activation, and direct audio stream access for real-time applications.
The library addresses the need for low-latency speech recognition in assistants, dictation tools, and streaming servers. It uses faster_whisper as its general-purpose default engine but supports multiple transcription backends through optional dependencies. The architecture distinguishes between streaming transcription for real-time text updates and final transcript refinement. For CPU-based production streaming on Linux x86-64, the tool recommends pairing sherpa-onnx-nemotron for fast incremental updates with sherpa-onnx-parakeet for authoritative final transcripts, processing only new audio frames during active speech turns rather than repeatedly retranscribing growing buffers. The library also includes native support for kroko-onnx as a local streaming ASR engine and silero-onnx-cpu for voice activity detection.
Developers should choose this tool if they need to integrate speech recognition into Python applications with minimal setup. It suits projects requiring microphone input handling, external audio stream processing, or browser-based streaming servers. The library is designed to work with only a few lines of code for basic use cases, while supporting advanced configurations for production deployments. The recommended engine profiles differ by hardware: CUDA setups continue using faster_whisper, while CPU deployments benefit from the two-engine approach that balances latency and quality.
The project maintains active development with regular updates to engine support and documentation. The maintainers have prioritized production deployment guidance through dedicated server setup documentation and pinned model configurations. Integration of community-contributed engines like kroko-onnx demonstrates responsiveness to user needs for specialized ASR solutions. The codebase targets Python 3.11 and 3.12 with documented platform-specific installation requirements, and the project uses multiprocessing for model operations with explicit guidance on script structure for cross-platform compatibility.
