Repository navigation
Official PyPI Downloads Verified
funasr-1.4.27-py3-none-any.whl: 1016211 bytes; SHA-2563341800380ed4e321a0103d25599c235a694a6ec4fbcb976d36ddbd948976fd7.funasr-1.4.27.tar.gz: 1092484 bytes; SHA-256a11075e1aa5591899ff5a2e2cae8d35d45e863f2d45423d3e059e61b93e87dd1.
Both official distributions match the GitHub Release bytes.
FunASR 1.4.27
Source: e227484d98ca90aa829a0ae180731c9ee2e87e77, tag v1.4.27.
Fixes
- Preserve valid NumPy int16 and multichannel audio-array inputs (#3890, #3891).
- Avoid dividing by zero for empty/zero-duration VAD speech segments (#3896).
- Preserve timestamp boundaries and handle omitted punctuation models (#3898).
- Accept supported segmentation-dictionary inputs in the character tokenizer (#3903).
These are root funasr package fixes. Standalone funasr-onnx and funasr-torch have separate versions and are not republished here. Open integer-PCM, UMAP, Paraformer-dictionary and sampler PRs are not included.
Verification Scope
- Release contracts: 20 passed after an 8-failure baseline. Release/docs/static checks: 133 passed, 2 skipped. Source and installed candidate behavior: 49 passed each; official 1.4.26 produced 15 expected failures among the same 49 contracts.
- Final-tag wheel and sdist were downloaded, matched to all packaged source bytes, checked with strict Twine and installed without network access. Final installed behavior: 49 passed, 0 skipped; 13 entrypoints checked.
- The exact four changed runtime source files remain bound to accepted CPU, real-L20Z and hosted-CI evidence. No new GPU run, pretrained-ASR accuracy, full-suite, Nano or production-upgrade claim.
- Exact-head, merged-main and final-tag CI logs, artifacts and checkout trees were accepted separately. Initial validation failures and subsequent fixes remain in the private maintenance record.
- Two Python payloads were newly downloaded. Ten unchanged runtime archives reused complete caches, freshly hashed against this release and the original runtime release, with archive integrity checks. They were not freshly downloaded or rebuilt. SHA256SUMS covers all 12 payloads; GitHub has 13 assets including that manifest.
- The prior normalization caveat remains: models trained with both norm_means=True and norm_vars=True need recognition-quality reevaluation; norm_vars=False remains the default.
PyPI publication and actual official-download verification are the separate final gate.
What's Changed
- Fix runtime hotword bounds on current main (preserves #3882) by @LauraGPT in #3885
- fix(runtime): cut long hotwords to 10 characters and skip empty ones by @tanishqbaweja in #3882
- Honor WebSocket offline model selection with CI (preserves #3887) by @LauraGPT in #3888
- fix(websocket): build the offline model from --asr_model by @tanishqbaweja in #3887
- fix(runtime): release ONNX hotword and timestamp fixes as 0.4.11 by @LauraGPT in #3889
- Fix WebSocket JSON hotword routing with CI (preserves #3880) by @LauraGPT in #3892
- fix(websocket): pass client hotwords to AutoModel as words by @tanishqbaweja in #3880
- fix(runtime): release Torch hotword and timestamp fixes as 0.1.7 by @LauraGPT in #3893
- docs(site): explain hotword protocols and standalone runtime upgrades by @LauraGPT in #3894
- docs(site): explain WebSocket offline model configuration by @LauraGPT in #3895
- Fix PCM16 and multichannel array input with NumPy CI by @LauraGPT in #3899
- fix(audio): average the channels of 2-D waveform arrays and tensors by @sohailshk in #3891
- fix(audio): scale int16 numpy waveforms to [-1, 1) in load_audio_text_image_video by @sohailshk in https://github.com//pull/3890
- docs(site): explain WAV input checks with executable PCM16 examples by @LauraGPT in #3901
- fix(auto): preserve zero-duration results and cover them in CI by @LauraGPT in #3902
- fix(auto): do not divide by zero speech time in AutoModel.inference by @sohailshk in #3896
- fix(timestamp): normalize sentence output without punctuation by @LauraGPT in #3904
- fix(timestamp): return string text from timestamp_sentence without punctuation by @kushaldave002 in #3898
- fix(tokenizer): skip blank segmentation entries with regression CI by @LauraGPT in #3905
- fix(tokenizer): skip blank lines when loading a seg_dict file by @MohammadHijjawi97 in #3903
- docs(site): explain dictionary input checks with executable UTF-8 examples by @LauraGPT in #3906
- docs(site): teach beginners a complete first-transcript workflow by @LauraGPT in #3914
- Release FunASR 1.4.27 by @LauraGPT in #3916
New Contributors
- @sohailshk made their first contribution in #3891
- @kushaldave002 made their first contribution in #3898
Full Changelog: v1.4.26...v1.4.27
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==1.4.27"