feat: Add SenseVoice as alternative STT backend — 5x faster, multilingual, emotion detection

[LauraGPT]

Motivation

Omi currently relies on Whisper/Deepgram for speech-to-text. SenseVoice (8.3K+ stars) offers significant advantages for a wearable AI device:

• 5x faster than Whisper large-v3 (non-autoregressive architecture, 234M params)
• 50+ languages with automatic language detection
• Built-in emotion detection (happy, sad, angry, neutral) — valuable for an AI companion
• Audio event detection (laughter, applause, music, crying) — enriches memory context
• Low latency — critical for real-time wearable experiences

Why this matters for Omi

1. Battery life: Faster inference = less compute = longer battery on wearable hardware
2. Richer context: Emotion + audio events give the AI companion deeper understanding of conversations
3. Self-hosted option: Users concerned about privacy can run SenseVoice locally (works on CPU)
4. Cost reduction: For cloud processing, faster inference means lower API costs

Integration

SenseVoice provides an OpenAI-compatible API server:

pip install funasr
funasr-server --device cuda

# Drop-in replacement for Whisper API
import openai
client = openai.OpenAI(base_url="http://localhost:8000/v1")
result = client.audio.transcriptions.create(
    model="iic/SenseVoiceSmall",
    file=open("audio.wav", "rb"),
)
print(result.text)  # transcription with emotion tags

Or use the Python API directly:

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input="audio.wav")
# Returns: text + language + emotion + audio events

Benchmarks

Model	Params	Speed (RTF)	CER (AISHELL-1)	Languages
Whisper-large-v3	1.5B	1.0x	8.4%	99
SenseVoice-Small	234M	5x	3.6%	50+

References

• GitHub: https://github.com/FunAudioLLM/SenseVoice
• Paper: FunAudioLLM: Voice Understanding and Generation Foundation Models
• PyPI: pip install funasr (1M+ monthly downloads)

Happy to help with integration if there's interest.