QuentinFuxa/WhisperLiveKit

WLK: Ultra-low-latency, self-hosted speech-to-text pipeline

• Simul-Whisper/Streaming (SOTA 2025) - Ultra-low latency transcription using AlignAtt policy.
• NLLW (2025), based on distilled NLLB (2022, 2024) - Simulatenous translation from & to 200 languages.
• WhisperStreaming (SOTA 2023) - Low latency transcription using LocalAgreement policy
• Streaming Sortformer (SOTA 2025) - Advanced real-time speaker diarization
• Qwen3-ASR-causal (2026) - Causal streaming audio encoder for Qwen3-ASR: each audio block is encoded exactly once, constant compute per audio second, append-only transcripts.
• AlignAtt4LLM (IWSLT 2026) - Simultaneous translation with decoder-only LLMs: attention-gated commits, append-only output.

Why not just run a simple Whisper model on every audio batch? Whisper is designed for complete utterances, not real-time chunks. Processing small segments loses context, cuts off words mid-syllable, and produces poor transcription. WhisperLiveKit uses state-of-the-art simultaneous speech research for intelligent buffering and incremental processing.

The backend supports multiple concurrent users. Voice Activity Detection reduces overhead when no voice is detected.

pip install whisperlivekit

# Start the server — open http://localhost:8000 and start talking
wlk --model base --language en


# Auto-pull model and start server
wlk run whisper:tiny

# Transcribe a file (no server needed)
wlk transcribe meeting.wav

# Generate subtitles
wlk transcribe --format srt podcast.mp3 -o podcast.srt

# Manage models
wlk models                             # See what's installed
wlk pull large-v3                      # Download a model
wlk rm large-v3                        # Delete a model

# Benchmark speed and accuracy
wlk bench

WhisperLiveKit exposes compatibility-oriented subsets of popular APIs:

# OpenAI-compatible REST API
curl http://localhost:8000/v1/audio/transcriptions -F file=@audio.wav

# Works with the OpenAI Python SDK
client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")

# Deepgram-compatible WebSocket
# See docs/API.md for supported options and current SDK setup

# Native WebSocket for real-time streaming
ws://localhost:8000/asr

Per-session WebSocket query parameters:

See docs/API.md for the complete API reference. For a native SwiftUI macOS client, see macos/WhisperLiveKitMac.

• See here for the list of all available languages.
• Check the troubleshooting guide for step-by-step fixes collected from recent GPU setup/env issues.
• For HTTPS requirements, see the Parameters section for SSL configuration options.

The Diart profile is limited to Python 3.11 and 3.12 because Diart 0.9.2 requires NumPy below 2. Use Sortformer for diarization on Python 3.13.

Supported GPU profiles:

# Profile A: Sortformer diarization
uv sync --extra cu129 --extra diarization-sortformer

# Profile B: Voxtral HF + translation
uv sync --extra cu129 --extra voxtral-hf --extra translation

# Profile C: Qwen3-ASR vLLM
uv sync --extra qwen3-vllm

qwen3-vllm uses vLLM's CUDA wheel stack and must be installed in a separate environment from cu129. Several heavy extras (voxtral-hf, qwen3-vllm-metal, and the vLLM stacks) intentionally conflict with one another and must be installed in separate environments; the authoritative list is [tool.uv].conflicts in pyproject.toml. The canary extra conflicts with voxtral-hf and qwen3-vllm-metal, but is compatible with diarization-sortformer (both pull nemo-toolkit[asr]).

See Parameters & Configuration below on how to use them.

Benchmarks use 6 minutes of public LibriVox audiobook recordings per language (30s + 60s + 120s + 180s), with ground truth from Project Gutenberg. The figures above were measured on an NVIDIA H100 (CUDA); the qwen3 causal tower is English-only, so it only appears on the English chart. Fully reproducible with python scripts/run_scatter_benchmark.py (the script picks a matching combo set on Apple Silicon: mlx-whisper and voxtral-mlx instead of the CUDA-only backends). Raw results: benchmarks/h100_scatter/. We are actively looking for benchmark results on other hardware (different NVIDIA GPUs, Apple Silicon chips, cloud instances). If you run the benchmarks on your machine, please share your results via an issue or PR!

Go to chrome-extension for instructions.

WhisperLiveKit supports Voxtral Mini, a 4B-parameter speech model from Mistral AI that natively handles 100+ languages with automatic language detection. Whisper also supports auto-detection (--language auto), but Voxtral's per-chunk detection is more reliable and does not bias towards English.

# Apple Silicon (native MLX, recommended)
pip install -e ".[voxtral-mlx]"
wlk --backend voxtral-mlx

# Linux/GPU (HuggingFace transformers)
pip install transformers torch
wlk --backend voxtral

Voxtral uses its own streaming policy and does not use LocalAgreement or SimulStreaming. See BENCHMARK.md for performance numbers.

Install the optional backend and run SenseVoiceSmall through WLK's existing LocalAgreement and VAC/VAD pipeline:

pip install "whisperlivekit[funasr]"
wlk --backend funasr --language auto

Use a verified local model snapshot without executing remote model code:

wlk --backend funasr --model_dir /path/to/SenseVoiceSmall --language yue

The initial integration supports SenseVoiceSmall transcription in Mandarin (zh), Cantonese (yue), English (en), Japanese (ja), Korean (ko), and automatic detection. FunASR uses LocalAgreement only; selecting it with the default policy switches that policy automatically. It does not support --direct-english-translation, and WLK remains responsible for voice activity control rather than enabling FunASR's internal VAD. This compatibility contract does not cover arbitrary FunASR models. SenseVoiceSmall is distributed under its model license.

qwen3-streaming runs Qwen3-ASR through plain HF Transformers with a bounded-recompute audio cache: the pretrained audio tower only re-encodes a local window (default 12 s) per update, cached audio embeddings are append-only, and text is committed with a stable-prefix rule. Works on CUDA, Apple Silicon (MPS) and CPU, no vLLM required.

pip install -e ".[qwen3-streaming]"
wlk --backend qwen3-streaming --language en

Notes:

• An explicit --language is required (automatic detection switches language mid-stream on accented audio).
• Word timestamps are interpolated estimates (~1 s precision): good enough for lines and diarization alignment. Use --backend qwen3-vllm (ForcedAligner) when exact word timing matters.
• Decode pacing self-adjusts to the hardware; on GPUs slower than real time the update cadence grows instead of lagging. Plan one realtime session per GPU.
• Defaults encode the validated operating point (12 s left context, ~15 s segments); see --help for the --qwen3-streaming-* knobs.

Causal mode (minimum compute per chunk). The windowed default re-encodes up to 12 s of audio on every update. The causal mode runs an append-only causal-KV encoder instead: each ~2 s audio block is encoded exactly once, memory is bounded (15 s window + sentence-boundary segment resets), and per-chunk compute is constant in stream length:

wlk --backend qwen3-streaming --language en \
    --qwen3-streaming-audio-backend causal \
    --qwen3-streaming-tower-checkpoint qfuxa/qwen3-asr-0.6b-streaming

The fine-tuned tower (qfuxa/qwen3-asr-0.6b-streaming) downloads automatically. The Qwen runtime, tests, experiments, benchmarks and figures now live in Qwen3-ASR-causal, which is consumed here through third_party/qwen3-asr-causal. WhisperLiveKit keeps only the backend wiring and CLI flags.

Use causal mode for many concurrent streams, energy-constrained serving or unbounded session lengths; keep the windowed default for best accuracy. English only for now. Detailed WER/RTF results are in the Qwen3-ASR-causal repository.

Experimental vLLM Metal causal mode. On Apple Silicon, qwen3-vllm-metal can also load the same causal tower and keep a rolling MLX decoder KV over the [prompt + audio] prefix. New audio blocks are encoded once, and the decoder replays only the new audio embeddings, prompt tail, and previous-hypothesis draft:

wlk --backend qwen3-vllm-metal --language en \
    --qwen3-vllm-metal-audio-backend causal \
    --qwen3-vllm-metal-tower-checkpoint qfuxa/qwen3-asr-0.6b-streaming

This path has passed local short-form smoke tests and is useful for comparing the Metal decoder path against the HF/MPS causal backend. It is still experimental; use qwen3-streaming causal for the validated production stack.

Experimental vLLM CUDA causal mode. On NVIDIA GPUs, qwen3-vllm can load the same causal tower, use vLLM for the text decoder, and keep vLLM ForcedAligner timestamps:

WLK_QWEN3_VLLM_LIVE_MULTIPROCESSING=1 \
wlk --backend qwen3-vllm --language en \
    --qwen3-vllm-audio-backend causal \
    --qwen3-vllm-causal-decoder-backend vllm-live \
    --qwen3-vllm-tower-checkpoint qfuxa/qwen3-asr-0.6b-streaming

The vllm-text backend is the conservative fallback when you do not want the live request-local append path; it still uses vLLM prefix caching but starts one text-decoder request per chunk. The append-kv and rolling names remain as compatibility aliases for the HF decoder path. Keep standard qwen3-vllm for best current accuracy until the causal quality gate is fixed.

WhisperLiveKit supports NVIDIA Canary-1b-v2 via NeMo, a 1B-parameter model covering 25 European languages with native word-level timestamps. Automatic language detection uses NeMo's AmberNet language-ID model when --language auto is set; the detected language is locked in once enough audio has accumulated. Canary streams through the LocalAgreement policy.

pip install -e ".[canary]"
wlk --backend canary --language auto

Notes:

• Runs on CUDA and CPU. A GPU is strongly recommended: CPU works (see scripts/smoke_canary.py) but is slow for a 1B-parameter model. On Apple Silicon, NeMo's current restore path uses CPU; this backend does not enable MPS device placement. NeMo is a heavy dependency (torch, Lightning, and friends).
• Supports WhisperLiveKit's Python range, 3.11 through 3.13. The canary extra intentionally selects nemo-toolkit[asr]>=3.0,<4, the NeMo line validated with Canary timestamps and Python 3.13.
• Explicit --language <code> skips language detection entirely. Tune detection with --canary-lid-min-sec (minimum audio before detecting) and --canary-lid-min-conf (confidence threshold to lock in the detected language).
• Canary-1b-v2 is distributed under CC-BY-4.0. The separate AmberNet language-ID model is downloaded from NVIDIA NGC; review its model-card terms for your deployment.

Command-line Interface: Start the transcription server with various options:

# Large model and translate from french to danish
wlk --model large-v3 --language fr --target-language da

# Diarization and server listening on */80
wlk --host 0.0.0.0 --port 80 --model medium --diarization --language fr

# Voxtral multilingual (auto-detects language)
wlk --backend voxtral-mlx

Python API Integration: Check basic_server for a more complete example of how to use the functions and classes.

import asyncio
from contextlib import asynccontextmanager

from fastapi import FastAPI, WebSocket, WebSocketDisconnect
from fastapi.responses import HTMLResponse

from whisperlivekit import AudioProcessor, TranscriptionEngine, parse_args

transcription_engine = None

@asynccontextmanager
async def lifespan(app: FastAPI):
    global transcription_engine
    transcription_engine = TranscriptionEngine(model_size="medium", diarization=True, lan="en")
    yield

app = FastAPI(lifespan=lifespan)

async def handle_websocket_results(websocket: WebSocket, results_generator):
    async for response in results_generator:
        await websocket.send_json(response)
    await websocket.send_json({"type": "ready_to_stop"})

@app.websocket("/asr")
async def websocket_endpoint(websocket: WebSocket):
    global transcription_engine

    # Create a new AudioProcessor for each connection, passing the shared engine
    audio_processor = AudioProcessor(transcription_engine=transcription_engine)    
    results_generator = await audio_processor.create_tasks()
    results_task = asyncio.create_task(handle_websocket_results(websocket, results_generator))
    await websocket.accept()
    while True:
        message = await websocket.receive_bytes()
        await audio_processor.process_audio(message)        

Frontend Implementation: The package includes an HTML/JavaScript implementation here. You can also import it using from whisperlivekit import get_inline_ui_html & page = get_inline_ui_html()

--sortformer-max-speakers is an assertion about the audio, not speaker-count estimation. Use it only when the session is known to contain at most N speakers. If more speakers are present, later arrival-ordered channels are excluded and those speakers can be attributed to a retained label. The default preserves the full checkpoint output. Sortformer produces independent activity probabilities, including overlapping activity, but WhisperLiveKit currently resolves each diarization frame and ASR token to one speaker. This option does not add simultaneous-speaker output.

For diarization using Diart, you need to accept user conditions here for the pyannote/segmentation model, here for the pyannote/segmentation-3.0 model and here for the pyannote/embedding model. Then, login to HuggingFace: huggingface-cli login

To deploy WhisperLiveKit in production:

1. Server Setup: Install production ASGI server & launch with multiple workers

   pip install uvicorn gunicorn
   gunicorn -k uvicorn.workers.UvicornWorker -w 4 your_app:app

2. Frontend: Host your customized version of the html example & ensure WebSocket connection points correctly

3. Nginx Configuration (recommended for production):

   server {
      listen 80;
      server_name your-domain.com;
       location / {
           proxy_pass http://localhost:8000;
           proxy_set_header Upgrade $http_upgrade;
           proxy_set_header Connection "upgrade";
           proxy_set_header Host $host;
   }}

4. HTTPS Support: For secure deployments, use "wss://" instead of "ws://" in WebSocket URL

Deploy the application easily using Docker with GPU or CPU support.

• Docker installed on your system
• For GPU support: NVIDIA Docker runtime installed

With GPU acceleration (recommended):

docker build -t wlk .
docker run --gpus all -p 8000:8000 --name wlk wlk

CPU only:

docker build -f Dockerfile.cpu -t wlk --build-arg EXTRAS="cpu" .
docker run -p 8000:8000 --name wlk wlk

Custom configuration:

# Example with custom model and language
docker run --gpus all -p 8000:8000 --name wlk wlk --model large-v3 --language fr

Compose (recommended for cache + token wiring):

# GPU Sortformer profile
docker compose up --build wlk-gpu-sortformer

# GPU Voxtral profile
docker compose up --build wlk-gpu-voxtral

# CPU service
docker compose up --build wlk-cpu

• Large models: Ensure your Docker runtime has sufficient memory allocated

• --build-arg Options:
  • EXTRAS="cu129,diarization-sortformer" - GPU Sortformer profile extras.
  • EXTRAS="cu129,voxtral-hf,translation" - GPU Voxtral profile extras.
  • EXTRAS="cpu,diarization-diart,translation" - CPU profile extras.
  • Hugging Face cache + token are configured in compose.yml using a named volume and HF_TKN_FILE (default: ./token).

# Quick benchmark with the CLI
wlk bench
wlk bench --backend faster-whisper --model large-v3
wlk bench --languages all --json results.json

# Install test dependencies for full suite
pip install -e ".[test]"

# Run unit tests (no model download required)
pytest tests/ -v

# Speed vs Accuracy scatter plot (all backends, compute-aware + unaware)
python scripts/create_long_samples.py        # generate ~90s test samples (cached)
python scripts/run_scatter_benchmark.py      # English (both modes)
python scripts/run_scatter_benchmark.py --lang fr  # French

Capture discussions in real-time for meeting transcription, help hearing-impaired users follow conversations through accessibility tools, transcribe podcasts or videos automatically for content creation, transcribe support calls with speaker identification for customer service...