FunASR-Nano ONNX

Environment

Python ≥ 3.8; install deps from requirements.txt.

Get models

• Download ready-made ONNX models:
  modelscope download --model zengshuishui/FunASR-nano-onnx --local_dir models

Export (only if you need to re-export from model.pt)

1. Edit scripts/run.sh and set your paths:

   • fun_asr_path=/path/to/Fun-ASR-Nano-2512

2. Run:

   cd scripts

   bash run.sh

   ONNX files will be placed in ../models/.

Decode demo

With downloaded or exported models, run:

bash decode.sh

decode.sh is a minimal demo, using INT8/FP32 models for quick sanity check.

Streaming ASR (WebSocket Server)

Run the real-time streaming ASR server:

cd streaming_fun_asr

python realtime_ws_server.py

Supports INT8 models (CPU) and FP32 models (CUDA), auto-detected. Open http://localhost:8000 for the demo page.

C++ inference

See detailed C++ examples in sherpa-onnx: https://github.com/k2-fsa/sherpa-onnx

Notes

• run.sh: one-click export (edit paths first).
• decode.sh: demo decode to verify models.
• Project Refactoring: The project has been refactored. The original main branch has been backed up to backup/main-2026-01-06.