FunASR v1.4.0: verified PyPI packages, safer AutoModel, and 12 release assets

[LauraGPT]

FunASR 1.4.0 is now available as a stable PyPI release:

python -m pip install -U "funasr==1.4.0"

What changed in the Python package:

• AutoModel rejects the common vda_model typo before downloading a model and points to vad_model.
• Installed wheels again include SenseVoice english.json, restoring English spelling normalization.
• Installed wheels include all four RWKV-BAT encoder/decoder .cpp and .cu sources required for runtime CUDA extension compilation.

The tagged GitHub source also finalizes legacy WebSocket file-input sessions after the end-control message. That runtime lives in the tagged source and is not part of the PyPI wheel.

The signed v1.4.0 GitHub Release contains 12 assets: the wheel, sdist, nine prebuilt llama.cpp / GGUF runtimes for Linux, macOS, and Windows, plus SHA256SUMS-v1.4.0.

The public PyPI wheel and sdist were downloaded again with no-cache headers and matched the exact-main build byte for byte. A fresh isolated install then passed metadata/import/version, the vda_model guard, SenseVoice English normalization, and all four RWKV-BAT kernel-source checks.

If you try 1.4.0 in a real application, please share the model, hardware, install path, input boundary, and any deployment friction. Reproducible reports help us keep the weekly release contract useful.