FunASR v1.4.5: Lighter Python Inference and Nine llama.cpp Runtimes

August 28, 2026 · Stable FunASR release

FunASR v1.4.5 makes the default Python inference install lighter: torchaudio is no longer a hard dependency, while kaldi-native-fbank is available when a native fbank path is needed. The same GitHub Release carries the wheel, sdist, nine llama.cpp / GGUF runtimes, and one checksum manifest.

Installation boundaries

Path	Command	Use it for
Standard	`python -m pip install -U "funasr==1.4.5"`	Normal Python inference without an automatic torchaudio install
KNF backend	`python -m pip install -U "funasr[knf]==1.4.5"`	A lighter feature path backed by kaldi-native-fbank
Existing torchaudio flow	`python -m pip install torchaudio`	Still supported when installed explicitly; it is outside the default dependency set

The stable distributions are on PyPI 1.4.5. Both were built from exact main commit 4df59cc1, downloaded again, and matched to their published digests.

Deployment evidence and limits

The recorded Ascend 910B run processed 70.47 seconds of audio in 1.15 seconds under its fixed test contract, for RTF 0.016. This is reproducible evidence for that hardware, model, audio, and timing scope, not a production capacity promise for other systems.

All 38 release-contract tests passed. The wheel and sdist passed Twine, archive integrity, isolated-install, and public-PyPI round-trip checks. The signed tag resolves exactly to 4df59cc15386ff3bb10916256d807ebc5c85f81d.

All 12 assets on one page

The v1.4.5 Release pins and mirrors runtime-llamacpp-v0.2.1 with:

Python: funasr-1.4.5-py3-none-any.whl and funasr-1.4.5.tar.gz
Linux: funasr-llamacpp-linux-arm64.tar.gz, funasr-llamacpp-linux-x64.tar.gz, funasr-llamacpp-linux-x64-avx2.tar.gz, funasr-llamacpp-linux-x64-vulkan.tar.gz
macOS: funasr-llamacpp-macos-arm64.tar.gz
Windows: funasr-llamacpp-windows-x64.zip, funasr-llamacpp-windows-x64-avx2.zip, funasr-llamacpp-windows-x64-cuda.zip, funasr-llamacpp-windows-x64-vulkan.zip
Verification: SHA256SUMS-v1.4.5

sha256sum -c SHA256SUMS-v1.4.5

Use the llama.cpp production deployment page for hardware selection, launch commands, smoke tests, and known boundaries. GPU API deployments can continue with the vLLM guide.

Start from a pinned version and verified assets, then revalidate on the target hardware with representative audio before rollout.

Open v1.4.5 and all 12 assets