Repository navigation
Official PyPI Downloads Verified
funasr-1.4.28-py3-none-any.whl: 1016346 bytes; SHA-256fce3969e68edd714a1d94bb4f186af845344f2aa176d4a9de6022fd0ff676823.funasr-1.4.28.tar.gz: 1094400 bytes; SHA-2561631cd4d2e1541a779753a26058431f0d291db804da3a23dfcbb67464e091de4.
Both official distributions match GitHub bytes and passed isolated installation and 45 behavior tests each.
FunASR 1.4.28
What Changed
- Whisper batches with different audio lengths: the frontend now uses each utterance's own length when calculating its features, rather than reusing the first item's length. See #3918.
- Symbols in timestamped text: core sentence postprocessing now preserves symbol-token timestamps and correctly handles following letter sequences. See #3929.
This updates the root funasr toolkit. It does not publish new standalone funasr-onnx or funasr-torch versions, change model weights, or upgrade a deployed service. The source-only WebSocket/ONNX changes since the previous tag are not claimed as additions to this wheel.
Verification and Limits
- Release version/news: 24 passed; release/docs/static checks: 133 passed, 2 browser-only skips. Source regressions: 45 passed; related checks: 35 passed. The complete text-postprocessing workflow selection also passed.
- Candidate and final-tag wheel/sdist were separately checked for archive integrity, exact packaged source, metadata, strict Twine, private offline installation, imports, CLI and 13 entrypoints. Each installed distribution passed 45 behavior tests. The official 1.4.27 baseline failed 18 of the same assertions.
- Fresh L20Z validation covered 100 real CUDA Whisper feature-extraction cases and six synthetic CUDA-generated token/postprocessor combinations. This is not a pretrained-model recognition-quality, accuracy or speed benchmark.
- A bounded repository-wide run remains incomplete: 83 passed, 2 skipped and 10 failed during online ModelScope lookup with networking disabled, before inference. It stopped at the configured maxfail boundary. No full-suite pass is claimed.
- Exact-head, merged-main and final-tag CI and their actual checkout trees were accepted separately. Initial environment/setup failures remain in the maintenance record.
- The two final Python artifacts were freshly downloaded. Ten unchanged runtime archives reuse complete cached bytes, freshly hashed against this release and the original runtime release and checked for archive integrity; they were not rebuilt or freshly downloaded.
SHA256SUMScovers all 12 payloads; the release has 13 assets including that manifest. - The existing normalization caveat still applies: checkpoints trained with both
norm_means=Trueandnorm_vars=Trueneed recognition-quality reevaluation;norm_vars=Falseremains the default.
PyPI publication and verification of actual official downloads are a separate final gate; the additional notice above appears only after that gate passes.
What's Changed
- docs(site): explain model selection through reader tasks by @LauraGPT in #3920
- fix(frontend): preserve per-item lengths in Whisper batches by @8079-debug in #3918
- docs(runtime): fix the dead Docker installation links in the SDK guides by @farisbasha in #3919
- docs(site): distinguish VAD intervals from audio edits and subtitles by @LauraGPT in #3925
- fix(websocket): send file-mode hotwords as the protocol's JSON object by @tanishqbaweja in #3924
- docs(site): explain recording-local speaker labels and transcript review by @LauraGPT in #3928
- fix(onnx): reset the frontend after a short final chunk in streaming Paraformer by @tanishqbaweja in #3927
- fix(postprocess): keep a timestamp for symbol tokens in sentence_postprocess by @sohailshk in #3929
- Release FunASR 1.4.28 by @LauraGPT in #3932
New Contributors
- @8079-debug made their first contribution in #3918
- @farisbasha made their first contribution in #3919
Full Changelog: v1.4.27...v1.4.28
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==1.4.28"