FunASR 1.4.3: optional Silero VAD and fixed-K speaker clustering / 可选 Silero VAD 与固定 K 聚类

[LauraGPT]

FunASR 1.4.3 is now available from public PyPI and the signed GitHub Release.

What changed

• AutoModel(vad_model="silero-vad") adds an optional Silero VAD adapter with FunASR-compatible millisecond segments, configurable thresholds, 8/16 kHz input, ONNX mode, and bounded segment lengths.
• Speaker diarization with a known speaker count now routes large embedding sets through fixed-K clustering instead of memory-heavy dense spectral clustering.
• Silero remains optional: existing FSMN VAD configurations and the base FunASR installation are unchanged.

python -m pip install -U "funasr==1.4.3"
# Add the optional Silero adapter:
python -m pip install -U "funasr[silero]==1.4.3"

from funasr import AutoModel

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")

Release verification passed 167 exact-main tests, compileall, PEP 517 build, Twine, source/artifact diff checks, and clean public-wheel installs. The GitHub Release contains exactly 12 assets: wheel, sdist, nine runtime-llamacpp-v0.2.0 platform archives, and SHA256SUMS-v1.4.3.

• PyPI: https://pypi.org/project/funasr/1.4.3/
• GitHub Release: https://github.com/modelscope/FunASR/releases/tag/v1.4.3
• English guide: https://www.funasr.com/en/blog/funasr-v1-4-3-pypi-release.html
• 中文指南: https://www.funasr.com/blog/funasr-v1-4-3-pypi-release.html

Public artifact SHA-256:

• wheel: 4b1491643a5bc6ccdd8acfc3ec438f9d9b9aaca8c2bf2d0e71f737dfc03025b9
• sdist: 74a9a60eac4f05b7cba25d31bdc0f6bba5d70dbe9fd326291b362a5e425ba01a

---

FunASR 1.4.3 已发布到 public PyPI 和带签名 tag 的 GitHub Release。

更新内容

• AutoModel(vad_model="silero-vad") 现可选用 Silero VAD 适配器，返回 FunASR 兼容的毫秒级片段，并支持阈值、8/16 kHz 输入、ONNX 模式和最长片段限制。
• 已知说话人数的说话人分离在大规模 embedding 输入下改用固定 K 聚类，避免内存开销较高的稠密谱聚类。
• Silero 仍是可选依赖：原有 FSMN VAD 配置和 FunASR 基础安装保持不变。

本次发布从 exact main 独立构建；167 项源码测试、compileall、PEP 517、Twine、源码/产物 diff 和 public wheel 隔离安装均通过。GitHub Release 恰好包含 12 个资产，并已从公网重新下载逐项核对大小与 SHA-256。

欢迎在反馈中附上模型、设备、采样率、输入边界和最小复现命令。