Feature: Add FunASR as self-hosted STT alternative to Deepgram

[LauraGPT]

Motivation

Omi currently uses Deepgram for speech-to-text, which requires API keys and incurs costs. FunASR (16K+ stars) provides a fully self-hosted, open-source alternative with an OpenAI-compatible API — meaning minimal integration effort.

Why FunASR

• Free & self-hosted: No API keys, no per-minute billing, data stays on your infrastructure
• OpenAI-compatible API: /v1/audio/transcriptions endpoint — drop-in replacement
• 50+ languages including English, Chinese, Japanese, Korean, etc.
• Industrial-grade accuracy: Paraformer (non-autoregressive, 170x realtime on GPU), SenseVoice (50+ languages, emotion detection)
• Built-in VAD + punctuation + speaker diarization (cam++)
• Runs on consumer GPUs: SenseVoice-Small (234M params) works on 4GB VRAM

Quick Start

pip install funasr vllm
funasr-server --device cuda  # starts OpenAI-compatible server at :8000

# Test
curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@audio.wav -F model=sensevoice

Since Omi already has Deepgram self-hosted deployment (Helm charts in backend/charts/), FunASR could serve as a lighter-weight, truly open-source alternative that's easier to deploy.

References

• GitHub: https://github.com/modelscope/FunASR
• Models: SenseVoice (50+ langs), Paraformer (zh/en), Fun-ASR-Nano (LLM-based)
• API docs: OpenAI-compatible /v1/audio/transcriptions