v1.4.24

Final Publication Acceptance

GitHub Release and PyPI 1.4.24 are both published. Both distributions were downloaded from official PyPI and match the accepted GitHub assets by SHA-256 and size:

• funasr-1.4.24-py3-none-any.whl: 1015161 bytes; SHA-256 c2678ebb77c7dfba1d85a79749ad30b441159b87d17eff901e0c6ad4ebcc06c7.
• funasr-1.4.24.tar.gz: 1081873 bytes; SHA-256 522dcb493be75aa57d8183e6197b4431c5e02fade689eda90cfd91f76e5edb57.

The official sdist download resumed after a transport timeout using an exact HTTP 206 range; the original partial and failure receipts were retained. The wheel was not downloaded again and PyPI was not re-uploaded.

FunASR 1.4.24

Source: 20f26b7c4905d010768ec91206149df2a90104b6 / tag v1.4.24.

Fix

Deliver #3852: ignore empty non-linguistic symbols instead of allowing the character/phoneme tokenizer scan to stall. Deliver #3853: correct variance normalization for padded utterances. Deliver #3854: preserve valid CIF output and gradients when a batch mixes utterances with zero and nonzero fires, including dynamic ONNX batches.

Compatibility: checkpoints trained with both norm_means=True and norm_vars=True require recognition-quality reevaluation after the corrected normalization. norm_vars=False remains the default. Local checkpoint round trips are not a pretrained recognition-quality evaluation.

This does not claim pretrained ASR quality, Windows/NPU acceptance or a full repository-suite result.

The separately released funasr-onnx==0.4.9 and funasr-torch==0.1.6 have their own tags, packages and verification. This root-package release does not republish them. PR #3848 is not included. This release does not claim every EOS path is fixed.

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 and final wheels: 37 passing CPU/ONNX cases each covering tokenizer symbols, variance normalization, local checkpoint round trips, repository-audio inference and CIF dynamic batches. The 27 CUDA parametrizations are intentionally skipped locally and separately covered by bound prior hardware evidence. All 439 package files and 13 entrypoints match the accepted source.
• Reused 24 tokenizer/CTC cases, 34 FP32 CIF cases and four normalizer repository-audio cases including six CUDA inference calls on L20Z, after rehashing 33,870 sealed evidence files and 18 helpers and binding the exact source to the installed wheel. The 106-case normalizer acceptance retains 102 unchanged cases and four repaired audio checks. Earlier harness/TF32-fixture failures remain preserved; they are not silently discarded or counted as passes. This release performs no new GPU execution and makes no pretrained-quality or full-Paraformer-export claim.
• The local candidate build and candidate/final package checks, tests and installations use private HOME/cache directories and separate network namespaces with verified network-unreachable probes. The final-tag build runs in GitHub Actions under the existing release workflow, not the local network namespace. No newly downloaded pretrained weights are used. The separately reported #3848 test-download incident is not an authorization or validation input for this release.
• Release-head CI was accepted from actual logs and recorded artifact inventories. The roadmap job checked out the pure head; the other head workflows used GitHub's temporary PR merge commit. Both checkout trees were verified against the exact candidate. Merged-main and final-tag source checkouts were separately verified. The main API-doc workflow also checked out only the retained Chinese API page from gh-pages into previous-pages; this separate documentation input was recorded, not treated as inference source. 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.
• No full repository-suite, production-environment upgrade, Nano benchmark or adoption/growth claim is made.

What's Changed

• fix(tokenizer): prevent empty-symbol hangs with hosted regression coverage by @LauraGPT in #3855
• fix(tokenizer): ignore empty non-linguistic symbols to avoid a hang by @MohammadHijjawi97 in #3852
• fix(normalize): correct variance with reproducible CPU/CUDA regressions by @LauraGPT in #3856
• fix(normalize): correct UtteranceMVN variance normalization by @BillVan1945 in #3853
• fix(cif): handle mixed zero-fire batches with CPU/ONNX regression CI by @LauraGPT in #3857
• fix(paraformer): handle trailing zero-fire samples in mixed CIF batches by @BillVan1945 in #3854
• Release FunASR 1.4.24 by @LauraGPT in #3858

Full Changelog: v1.4.23...v1.4.24

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==1.4.24"