v1.4.20

Final Publication Acceptance

GitHub Release and PyPI 1.4.20 are both published. Both distributions were downloaded from official PyPI and match the accepted GitHub assets by SHA-256 and size:

• funasr-1.4.20-py3-none-any.whl: 1014313 bytes; SHA-256 664a1fbf36bc177a5aec3230b2b1023c160daf544da540bea60c7f3acd6393c3.
• funasr-1.4.20.tar.gz: 1063112 bytes; SHA-256 906f1a496acbc7235b8739cafd487a8e6d47b9785f466c37c445590ad4531a85.

FunASR 1.4.20

Source: f122f6c7bc6176ed307cc1659b2bcc031d54791f / tag v1.4.20.

Fix

Deliver #3819 (integrating #3788): optional speaker imports no longer disappear behind an unrelated NameError. When spk_model is requested and an optional speaker dependency failed to import, the constructor reports an actionable ImportError chained to that original failure. ASR-only construction remains available.

This changes error reporting, not the underlying missing/incompatible dependency on the reporter's Windows machine. It does not claim Windows hardware acceptance, pretrained ASR/diarization quality or resolution of every speaker issue.

The separately released funasr-onnx==0.4.5 has its own tag, packages and verification. This root-package release does not republish it.

Verification Scope

• New candidate release/README checks: 45 cases, 0 skips. Related docs/package/release checks: 104 cases, 2 explicitly recorded skips; no failures or errors.
• Actual installed candidate wheel: 10 fresh-process CPU optional-import cases, including both kwargs/config entry points, missing-module/runtime-initialization errors and healthy controls. Controlled model constructors are not pretrained ASR acceptance. All 439 package files and 13 entrypoints match the accepted source.
• Reused #3819's five real L20Z FP32 synthetic constructor/import-boundary results by rehashing all 172 evidence entries and binding all four recorded source modules to the installed wheel. Those cases include native reduced SANM/CIF CUDA inference and healthy native CAM++ embeddings, not new GPU execution, BF16, Windows or pretrained-quality validation.
• Exact release-head and merged-main CI were accepted with actual logs/artifacts. The final-tag workflow and the newly downloaded Python wheel/sdist passed byte/metadata/archive/source checks, strict Twine, offline installation, actual import and CLI checks.
• The 10 unchanged runtime archives reuse complete cache bytes freshly hashed against new asset metadata and the existing runtime source release. They were not rebuilt or re-executed. SHA256SUMS covers all 12 payloads. This is not a new runtime-hardware claim.
• The historical #3819 native BLAS performance diagnostic failure and later private-environment repair remain separate records. No full repository-suite, production-environment upgrade, Nano benchmark or adoption/growth claim is made.

What's Changed

• fix(examples): avoid smoke-test collection name collisions by @LauraGPT in #3812
• fix(runtime): repair evaluation CLIs and honor GPU device selection by @LauraGPT in #3813
• fix(runtime): include Torch runtime sources and required dependencies in distributions by @LauraGPT in #3814
• fix(examples): keep standalone client checks out of pytest fixture discovery by @LauraGPT in #3815
• docs: disclose unmerged OpenClaw SDK dependency by @LauraGPT in #3816
• fix(runtime): release the repaired Torch package as 0.1.4 by @LauraGPT in #3817
• docs: reconcile LocalAI manual integration status by @LauraGPT in #3818
• fix(auto): preserve optional speaker dependency errors by @LauraGPT in #3819
• docs: scope Ascend compatibility and latency evidence by @LauraGPT in #3820
• fix(onnx): handle short chunks and short streams in Fsmn_vad_online by @tanishqbaweja in #3822
• fix(runtime): release streaming ONNX wrapper fixes as 0.4.4 by @LauraGPT in #3826
• fix(runtime): handle streaming Paraformer chunks without features by @LauraGPT in #3827
• fix(runtime): release streaming ONNX wrapper fixes as 0.4.5 by @LauraGPT in #3828
• Release FunASR 1.4.20 by @LauraGPT in #3829

New Contributors

• @tanishqbaweja made their first contribution in #3822

Full Changelog: v1.4.19...v1.4.20

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.2.6.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Platform	Asset	SHA-256
Linux arm64	funasr-llamacpp-linux-arm64.tar.gz	7bca29cfa3c9a08e235a62212ca9e00f6656e59a8f07078966a2bfda1e5aa1f9
Linux x64 AVX2	funasr-llamacpp-linux-x64-avx2.tar.gz	aaebc5470f846ce915200b35d6e9f9bd0a0d3ed399d39e49bdeb7a1f1782bc70
Linux x64 Vulkan	funasr-llamacpp-linux-x64-vulkan.tar.gz	f02d41e98e9d4041f0896661007193810f025484d2175958f7c1313d5c90ec46
Linux x64 portable	funasr-llamacpp-linux-x64.tar.gz	779967de1c528c2be966bcc47f246e7d3e6fcdb748d9491263062f4120f35e52
macOS arm64	funasr-llamacpp-macos-arm64.tar.gz	bda59474202b887190f59d25b7b42c714469efae71276072c12fa0a38de68792
Windows x64 AVX2	funasr-llamacpp-windows-x64-avx2.zip	062cda8fefadd31c3e811227116daccf448a8520f4b0bb168d225c896e65ebbd
Windows x64 CUDA Blackwell (sm_120)	funasr-llamacpp-windows-x64-cuda-blackwell.zip	e32961a753f40888182f352fa551159c5165a6a77718ae4ade316aedfea4b1c2
Windows x64 CUDA	funasr-llamacpp-windows-x64-cuda.zip	148657911fb666b7af6ec43af2e23a0984e3259012b4c39f95631b717feb6840
Windows x64 Vulkan	funasr-llamacpp-windows-x64-vulkan.zip	debf8007e55011cad06081e7b8a78972f1b8fe672bc324d41e650d68821f6a6a
Windows x64 portable	funasr-llamacpp-windows-x64.zip	f6a73a548413ba9fbaf2145263ea66ec53cbdad1fb11790dbeeee493e339492e

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.4.20"