FunASR v1.4.0: Complete PyPI Packages and Safer AutoModel Arguments

July 31, 2026 · FunASR weekly stable release

FunASR v1.4.0 turns “present in source” into a tested installation contract. The PyPI wheel restores the SenseVoice English spelling table and RWKV-BAT CUDA build sources, while rejecting the common vda_model typo before model download and pointing to vad_model.

The signed GitHub Release contains the wheel, sdist, nine self-contained llama.cpp / GGUF runtimes, and SHA256SUMS-v1.4.0: 12 assets in total.

Python fixes and distribution boundaries

Distribution boundary	v1.4.0 content	Why it matters
PyPI wheel	`vda_model` fails before model download and points to `vad_model`	VAD segmentation, speaker processing, and `sentence_info` are no longer silently disabled by a typo
PyPI wheel	SenseVoice `english.json` is installed	British spellings such as `colour centre` reliably normalize to American spellings
PyPI wheel	Four RWKV-BAT encoder/decoder `.cpp`/`.cu` files are installed	Runtime CUDA extension compilation can locate its sources after wheel installation
tagged source	The legacy WebSocket file client/server completes offline, online, and 2pass tail audio after end-control and returns an explicit acknowledgement	This fix ships in the GitHub tagged source and is not part of the PyPI wheel; clients no longer mistake a fixed sleep for completion

1. Install and confirm the version

Pin the stable build from public PyPI:

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

Public PyPI contains exactly two Python files: funasr-1.4.0-py3-none-any.whl and funasr-1.4.0.tar.gz. Their sizes and SHA-256 values match the exact-main build, and both were downloaded with no-cache headers before a second isolated installation smoke.

The release gate passed 87 package/AutoModel/WebSocket/docs/runtime/llama.cpp tests (with one known online model-download test deselected for the host environment) plus 52 product-site tests. Twine 7.0 accepted both artifacts; the public wheel then passed version, import, typo guard, SenseVoice normalization, and four-file RWKV-BAT kernel checks.

2. Nine runtime targets and one checksum manifest

The same release page reuses and verifies every asset from runtime-llamacpp-v0.1.9:

linux-arm64, linux-x64, linux-x64-avx2, and linux-x64-vulkan
macos-arm64
windows-x64, windows-x64-avx2, windows-x64-cuda, and windows-x64-vulkan

Verify the download directory before use:

sha256sum -c SHA256SUMS-v1.4.0

See the llama.cpp / GGUF page for model and hardware selection. The manifest covers the wheel, sdist, and nine runtimes, while GitHub also exposes a digest for the manifest itself.

Pin the Python version, verify downloaded assets, and reproduce with real audio. When reporting a problem, include the model, device, input boundary, and exact command.

Open v1.4.0 and all 12 assets