Add Fun-ASR-Nano (Speech LLM) to the list

[LauraGPT]

Suggestion

This awesome list of multimodal LLMs should include Fun-ASR-Nano — a speech LLM that combines audio understanding with language generation.

Fun-ASR-Nano

• Architecture: SenseVoice encoder + Qwen3-0.6B decoder
• Parameters: 800M total
• Languages: 31 languages, dialects, accents
• Task: End-to-end speech recognition with LLM-based decoding
• GitHub: https://github.com/FunAudioLLM/Fun-ASR (1.2K stars)

SenseVoice (Audio Encoder)

• Architecture: Non-autoregressive audio encoder
• Capabilities: Speech recognition + emotion detection + audio event classification
• Parameters: 234M
• Languages: 50+
• Paper: https://arxiv.org/abs/2407.04051
• GitHub: https://github.com/FunAudioLLM/SenseVoice (8.3K stars)

Why relevant to this list

Fun-ASR-Nano represents the "audio modality + LLM" paradigm — using a pre-trained audio encoder (SenseVoice) connected to a language model decoder (Qwen3), similar to how vision-language models connect CLIP/SigLIP to LLMs. This is a key architecture pattern in the multimodal LLM space.

The broader FunASR ecosystem: https://github.com/modelscope/FunASR (16.6K stars)