Voice Transcription Proxy

OpenAI-compatible transcription service powered by LFM2.5-Audio-1.5B (Liquid AI).

Uses the official liquid-audio Python package with PyTorch for native GPU inference — no llama.cpp needed.

Endpoints

Endpoint	Protocol	Description
POST /v1/audio/transcriptions	HTTP	Whisper-compatible file upload
POST /v1/audio/transcriptions + stream=true	HTTP SSE	Streaming transcription
ws:///v1/realtime?intent=transcription	WebSocket	OpenAI Realtime API with server VAD
GET /health	HTTP	Health + GPU status

Run locally

uv sync
uv run python transcription_proxy.py --port 8091

Docker

docker build -t voice .
docker run --gpus all -p 8091:8091 voice

# With persistent model cache
docker run --gpus all -p 8091:8091 -v voice-cache:/cache voice

Kubernetes

containers:
  - name: voice
    image: ghcr.io/anthaathi/voice-transcription-proxy:latest
    ports:
      - containerPort: 8091
    resources:
      limits:
        nvidia.com/gpu: 1
    volumeMounts:
      - name: cache
        mountPath: /cache

Architecture

┌─────────────────────────────────────────┐
│ Single process (FastAPI + liquid-audio)  │
│                                          │
│  Model loaded on GPU at startup          │
│  ├─ POST /v1/audio/transcriptions        │
│  ├─ WebSocket /v1/realtime (webrtcvad)   │
│  └─ ASR via LFM2AudioModel              │
└─────────────────────────────────────────┘

Model is auto-downloaded from HuggingFace on first start. Mount /cache volume to persist across restarts.