Real-time transcription comes with an awkward tradeoff: cloud APIs are accurate but add network latency and ship your voice off-device, while most local options either lag or sacrifice quality. This library treats that as an orchestration problem rather than a model problem — it layers voice-activity detection, wake-word gating, and a swappable transcription engine so the expensive Whisper pass only runs on speech that actually matters, keeping perceived latency under a second on a local GPU.
What Sets It Apart
- Two-stage voice-activity detection (cheap WebRTC gating in front of accurate Silero) keeps the GPU from transcribing silence — lower compute and faster turnaround for always-listening apps.
- The transcription engine is pluggable rather than hardcoded: faster-whisper by default, with whisper.cpp, OpenAI Whisper, Moonshine, sherpa-onnx, and Kroko-ONNX as drop-ins, so you can trade accuracy for speed without rewriting your pipeline.
- Wake-word activation (Porcupine or OpenWakeWord) is built in, so it works as the front end of a voice assistant, not just a dictation box.
- It emits both live partial hypotheses and a final corrected transcript, so a UI can stream text as you speak yet still settle on an accurate result.
Who It's For
Great fit if you're building a voice assistant, dictation tool, or browser-based streaming server and want it all to run locally with low latency and no per-minute API bill. Look elsewhere if you need a turnkey hosted endpoint, broad mobile support, or aren't ready to manage Python 3.11+, PortAudio, and CUDA — the flexibility comes at the cost of assembling and tuning the stack yourself.
