FunASR v1.4.3: Optional Silero VAD and Fixed-K Speaker Clustering

August 21, 2026 · FunASR weekly stable release

FunASR v1.4.3 adds an optional adapter for AutoModel(vad_model="silero-vad"). It returns FunASR-compatible millisecond segments and supports threshold controls, 8/16 kHz input, ONNX mode, and bounded segment lengths.

Speaker diarization with a known speaker count now sends large embedding sets through fixed-K clustering instead of memory-heavy dense spectral clustering. The signed GitHub Release also contains the wheel, sdist, nine self-contained llama.cpp / GGUF runtimes, and SHA256SUMS-v1.4.3: 12 assets in total.

What changed

Area	v1.4.3 behavior	Why it matters
Optional VAD	`silero-vad` and `silero_vad` aliases use the existing AutoModel VAD pipeline	Choose Silero without changing downstream ASR code
Input and output	8/16 kHz waveforms produce millisecond ranges, with threshold, minimum-silence, ONNX, and maximum-segment controls	Segments can feed FunASR long-audio recognition directly
Speaker clustering	Large embedding sets with a known speaker count use fixed-K clustering	Avoids the quadratic memory pressure of dense spectral clustering
Compatibility	Silero remains an optional `funasr[silero]` dependency; the existing FSMN VAD path is unchanged	Base installs and established production configurations remain stable

1. Install and confirm the version

Pin the stable build from public PyPI:

python -m pip install -U "funasr==1.4.3"
funasr --version

Enable Silero VAD:

python -m pip install -U "funasr[silero]==1.4.3"

model = AutoModel(
    model="paraformer-zh",
    vad_model="silero-vad",
    device="cpu",
    vad_kwargs={"silero_threshold": 0.5, "silero_min_silence_duration_ms": 100},
)
result = model.generate(input="audio.wav")

Public PyPI contains exactly funasr-1.4.3-py3-none-any.whl and funasr-1.4.3.tar.gz. Both were built independently from exact main and downloaded again with no-cache headers for isolated installation smokes.

All 167 exact-main source tests passed, followed by compileall, PEP 517 build, Twine, source/artifact diff checks, and isolated installs with both no dependencies and full dependencies. The public wheel, sdist, and all 12 GitHub assets were downloaded again and matched by size and SHA-256.

2. Nine runtime targets and one checksum manifest

The same release page reuses and verifies every asset from runtime-llamacpp-v0.2.0:

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-v1.4.3

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 v1.4.3 and all 12 assets