WhisperLiveKit

Use it when

• You need live low-latency captions with speaker labels from meetings or streams without sending audio to a cloud service
• You want a self-hosted drop-in replacement for OpenAI or Deepgram transcription endpoints that your existing SDK clients can point at

What it solves

Not the fit when

• Qwen3 causal streaming is English-only
• Diarization is not included in the base installation; install either the Sortformer or Diart diarization extra
• Diart requires accepting the referenced pyannote model conditions and logging in to Hugging Face
• Several optional backend stacks are intentionally incompatible and must be installed in separate environments
• The server defaults to localhost and supports optional API-token authentication; when exposing it beyond localhost, configure --api-token, restricted CORS, TLS or a TLS reverse proxy, and upload/resource limits
• Large models and advanced backends require hardware and memory appropriate to the chosen model; the retained H100 benchmark is not a performance guarantee for other hardware
• text to speech generation
• managed cloud transcription at scale
• audio translation into spoken output
• non-speech audio analysis

Install

pip install whisperlivekit for the basic server and file-transcription CLI; optional MLX, CUDA, translation, Qwen, Voxtral, and diarization backends require the corresponding extras documented in the source checkout, and some backend combinations require separate environments

Invoke

Run wlk --model base --language en and open http://localhost:8000; use wlk transcribe meeting.wav for files, POST audio to http://localhost:8000/v1/audio/transcriptions for the OpenAI-compatible REST endpoint, or connect to ws://localhost:8000/asr for native streaming. Speaker labels require installing a diarization extra before adding --diarization.

Alternatives

No reviewed alternatives recorded yet.