Related work: FunASR SenseVoice — multi-task speech model with emotion recognition

[LauraGPT]

Hi! Interesting open-source speech language model.

Wanted to share a related project that might be of interest for comparison or collaboration:

SenseVoice

SenseVoice is a multi-task speech foundation model that handles:

• ASR (50+ languages)
• Speech emotion recognition (happy, sad, angry, neutral, etc.)
• Audio event detection (laughter, applause, music, etc.)

All in a single model with ultra-low latency (~70ms for 10s audio).

Comparison areas

	Tada	SenseVoice
Architecture	Speech LM	Multi-task encoder
Tasks	Speech understanding	ASR + emotion + events
Latency	-	~70ms/10s
Languages	-	50+

Also relevant

• Fun-ASR-Nano — Encoder + LLM speech model (audio encoder + Qwen2.5-0.5B): https://github.com/FunAudioLLM/Fun-ASR
• FunASR — Full ASR toolkit: https://github.com/modelscope/FunASR (16K+ stars)

Paper: https://arxiv.org/abs/2407.04051