v1.4.17

FunASR 1.4.17

Release source: 19e33badf1dba6f7606c9410c32c622486b5df13 (tag v1.4.17).

• 487 focused pytest checks passed; two local browser cases were skipped. The final PR-head hosted browser and other CI checks passed separately.
• Real NVIDIA L20Z validation includes CUDA compile checks and 48 CPU/CUDA FP32/BF16 keyword-decoding comparisons. These use synthetic inputs or tiny stand-ins, not pretrained ASR quality evaluation.
• Published wheel and sdist source bytes were checked against the final tag; strict Twine, offline installation, import and CLI checks passed.
• All ten runtime archives (including CUDA Blackwell), wheel and sdist were byte/hash and integrity verified. Eight files were downloaded from this release; four unchanged runtime archives reuse complete verified cache bytes whose full SHA-256 and size equal the new GitHub asset metadata. The new Python packages do not reuse candidate builds. SHA256SUMS covers these 12 payloads. Archive integrity was checked; platform binaries were not executed.

This is focused release validation, not the complete repository test suite or the Nano leaderboard regression.

What's Changed

• Publish FunASR 1.4.16 bilingual site guide by @LauraGPT in #3721
• Record FunASR 1.4.16 release closeout by @LauraGPT in #3722
• fix(campplus): bound the clustering thread usage and stop over-computing the spectrum by @Gmgge in #3699
• fix(paraformer): correct timestamp alpha and peak argument order by @xx205 in #3716
• [perf] Fun-ASR-Nano fine-tuning on B200: 0.244 s to 0.068 s per step at 1 GPU, mostly from cuDNN SDPA plan builds on every new batch shape by @TarzanZhao in #3705
• fix: use y.device instead of hardcoded .cuda() in data2vec compute_var by @li-lizhe in #3710
• docs: note mixed-device support in PyPI 1.4.16 by @LauraGPT in #3724
• docs(nano): clarify training-speed evidence and release scope by @LauraGPT in #3725
• docs: 新增《在昇腾 NPU 上运行 FunASR》指南 by @Liuzi0816 in #3708
• fix(nano): correct punctuation timestamps across VAD merges by @Bruce-Yii in #3703
• fix: avoid deep-copying the checkpoint in load_pretrained_model by @yydhYYDH in #3729
• Remove unused imports by @yaojingguo in #3731
• Fix invalid escape sequences in text normalization and models by @MBK-fr in #3709
• fix: assign sentence speakers by total overlap without double counting by @LauraGPT in #3733
• feat(benchmark): report first nonempty transcript latency by @LauraGPT in #3734
• fix(sensevoice): preserve SentencePiece word boundaries by @LauraGPT in #3737
• fix(website): correct LiteLLM transcription route and guards by @LauraGPT in #3736
• docs: diagnose actual AutoModel GPU placement by @LauraGPT in #3740
• fix(amp): resolve autocast/GradScaler device type for non-CUDA accelerators by @li-lizhe in #3735
• fix(auto): warn when unavailable accelerators fall back to CPU by @LauraGPT in #3742
• Fix stale start timestamp in timestamp_sentence_en when timestamp list is shorter than punctuation list by @Arthur031221 in #3741
• docs: surface merged speech-to-speech SenseVoice integration by @LauraGPT in #3743
• docs(ops): record verified product-site release by @LauraGPT in #3745
• fix(hotwords): stop fuzzy postprocess matching from rewriting nearby text by @Arthur031221 in #3744
• fix: keep token separators in timestamp_sentence raw_text by @MohammadHijjawi97 in #3746
• fix(sensevoice): stop rendering the Cough event as the Sneeze emoji by @MohammadHijjawi97 in #3747
• fix(runtime): skip empty deduplicated event segments by @LauraGPT in #3748
• ci: run rich transcription regressions in ONNX package checks by @LauraGPT in #3749
• feat(audio): support explicit PCM16 input without container probing by @LauraGPT in #3751
• ci: cover CLI regressions in CPU NumPy checks by @LauraGPT in #3752
• docs(ops): make site monitoring aware of log rotation by @LauraGPT in #3753
• fix(realtime): recover merged long-segment partials by @LauraGPT in #3755
• fix(auto-model): preserve the resampled rate for speaker inference by @hulkbig in #3763
• fix: retain unpunctuated tail in sentence timestamps by @Arthur031221 in #3754
• fix(auto): skip blank lines in wav.scp and jsonl file lists by @Lesereingrape in #3758
• fix(paraformer): skip blank lines and read hotword files as UTF-8 by @MohammadHijjawi97 in #3761
• ci: cover timestamp and Paraformer hotword regressions by @LauraGPT in #3765
• docs: fix contact image paths in runtime SDK tutorials by @pratikgx in #3764
• fix(frontend): stop cmvn readers from crashing on blank lines or returning empty statistics by @Lesereingrape in #3760
• fix(auto): generate independent keys for unkeyed batch inputs by @hulkbig in #3772
• fix(data): read and write scp2jsonl/jsonl2scp lists as UTF-8 by @MohammadHijjawi97 in #3774
• docs: fix SenseVoice README images and Android SDK guide link by @pratikgx in #3775
• fix(train): bind distributed training to the active accelerator by @li-lizhe in #3771
• fix(metrics): score utterances with an empty reference in funasr.metrics.wer by @MohammadHijjawi97 in #3779
• fix(nano): avoid duplicate fs and preserve audio loading errors by @txfs19260817 in #3778
• fix(metrics): size the WER cost matrix for long inputs by @Batianxieshen1 in #3784
• fix(metrics): size the duplicated WER cost matrices for long inputs by @Batianxieshen1 in #3785
• fix(metrics): report unscored references in the WER summary line by @MohammadHijjawi97 in #3787
• fix(cli): warn when jieba is unavailable for readable SRT splitting by @Linxiushen in #3726
• fix(kws): ignore empty keyword entries in KwsCtcPrefixDecoder by @MohammadHijjawi97 in #3800
• Release FunASR 1.4.17 by @LauraGPT in #3770

New Contributors

• @Gmgge made their first contribution in #3699
• @xx205 made their first contribution in #3716
• @TarzanZhao made their first contribution in #3705
• @Liuzi0816 made their first contribution in #3708
• @Bruce-Yii made their first contribution in #3703
• @yydhYYDH made their first contribution in #3729
• @yaojingguo made their first contribution in #3731
• @MBK-fr made their first contribution in #3709
• @Arthur031221 made their first contribution in #3741
• @MohammadHijjawi97 made their first contribution in #3746
• @hulkbig made their first contribution in #3763
• @Lesereingrape made their first contribution in #3758
• @pratikgx made their first contribution in #3764
• @txfs19260817 made their first contribution in #3778
• @Batianxieshen1 made their first contribution in #3784
• @Linxiushen made their first contribution in #3726

Full Changelog: v1.4.16...v1.4.17

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...

FunASR 1.4.16

FunASR 1.4.16

Highlights

• Reliable mixed-device pipelines: explicit vad_kwargs, punc_kwargs, and spk_kwargs device placement is preserved across repeated inference calls. Runtime device overrides no longer mutate the placement of already-loaded models, caller-owned submodel configuration stays unchanged, and speaker inference uses the speaker model's own configuration.
• Clearer punctuation behavior: documentation now explains when SenseVoiceSmall or Paraformer output needs the CT-Transformer punctuation postprocessor instead of assuming every model emits punctuation natively.
• Native Transformers adoption guides: new tested English and Chinese guides connect model discovery, version requirements, CUDA installation, local downloads, and bounded acceptance checks.

Install

python -m pip install -U "funasr==1.4.16"

Install a PyTorch/TorchAudio build appropriate for your hardware and the selected model's optional dependencies.

Verification boundaries

• Mixed-device regressions were exercised with CPU-resident stand-in weights and mocked accelerator availability. They verify configuration propagation, fallback, repeated calls, and model/config consistency; they are not a CUDA, MPS, or NPU performance certification.
• Native Transformers guidance is versioned source documentation. Follow each model's documented revision and hardware requirements; it does not claim that every model/backend combination is interchangeable.
• Native downloads retain their independently versioned runtime-llamacpp-v0.2.6 source commit a57c05bfe2a91b5e0cb0983479634eba3e28ede5. They are unchanged native binaries, not builds from the Python 1.4.16 commit and not carriers of Python-only fixes. The separate Blackwell archive remains on the runtime release and is not attached here.

Previous Python release: 1.4.15. For a regression, retain the last working environment and report the exact artifact, model revision, platform, and reproduction.

Full source changes

Runtime downloads

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

The same nine verified runtime assets are attached directly to this Python release.

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	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

For Python users, install from PyPI:

python -m pip install -U "funasr==1.4.16"

Artifact verification

SHA256SUMS covers the wheel, sdist, and nine attached native archives. The Python artifacts are:

• funasr-1.4.16-py3-none-any.whl: f95943f6a8111349a1b08907363b048a49e9cb9caa1e3afd8c92970d08bf450d
• funasr-1.4.16.tar.gz: 1e2f8e3887fbc4b164cce40f73bf1324907da6ab59d1e8bf2add827f534c69e7

PyPI and GitHub Python artifact bytes must match after publication.

v1.4.15

FunASR 1.4.15

Highlights

• NumPy 2 compatibility: remove the NumPy upper bound with real frontend, masking and speaker-clustering tests; update EEND mel-filter calls for modern librosa. Thanks to @linhongyu510 for the original compatibility contribution. (#3674)
• Streaming and training reliability: preserve KWS frames across packet/EOS boundaries and honor optional output directories; exclude idle time from the dynamic VAD silence schedule; exclude unavailable validation metrics from checkpoint ranking. (#3655, #3656, #3676, #3677)
• MOSS and deployment documentation: add explicit MOSS profiles to the Gradio client and an offline vLLM example for whole-recording transcription with timestamps and anonymous speaker labels. Refresh task-oriented multilingual documentation and deployment entry points. (#3669, #3678)

Install

python -m pip install -U "funasr==1.4.15"

Install a PyTorch/TorchAudio build appropriate for your hardware and the selected model's optional dependencies. Backend environments remain separate: the default Qwen3 example environment is not interchangeable with the MOSS backend environment.

The MOSS offline example is a source recipe, not a newly installed wheel command. Follow the versioned MOSS guide for model/backend requirements. MOSS is an OpenMOSS model; speaker labels distinguish speakers within a recording and do not identify a known person.

Verification Boundaries

• NumPy compatibility was tested on Linux/Python 3.12 with NumPy 1.26.4 and 2.4.0, Torch/TorchAudio 2.10.0 CPU, and the recorded scientific-stack dependencies. This is not a guarantee for every historical dependency, Python version, GPU or optional model combination.
• Streaming and checkpoint regression coverage does not establish that every microphone tail-loss, model-accuracy or continual-learning issue is resolved. Reporter issues remain open for published-package feedback.
• MOSS model-smoke evidence uses a fixed model revision and backend. It does not establish complete long-recording tail coverage, multi-speaker accuracy, CER/WER improvement or a new performance benchmark.
• Native downloads retain their independently versioned runtime-llamacpp-v0.2.6 source commit a57c05bfe2a91b5e0cb0983479634eba3e28ede5. They are unchanged native binaries, not builds from the Python 1.4.15 commit and not carriers of Python-only fixes. Cross-platform build success is not a hardware-inference certification; CUDA/Blackwell architecture and driver requirements still apply.

For a regression, retain the last working environment and report the exact artifact, model revision, platform and reproduction. Previous Python release: 1.4.14. Downgrading FunASR alone does not restore other changed dependencies.

Full source changes

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==1.4.15"

Artifact verification

SHA256SUMS covers the wheel, sdist and ten native archives. PROVENANCE.json records their independent sources and verification scope. PyPI and GitHub Python artifact bytes match.

funasr-onnx 0.4.3

Standalone ONNX Runtime Python package

Install with:

python -m pip install --upgrade "funasr-onnx==0.4.3"

This is the funasr-onnx distribution, not a new release of the separate funasr toolkit. Updating funasr alone does not update these wrappers.

Fixes delivered

• Preserve missing optional exporter dependencies as ImportError causes instead of masking them with invalid string raises.
• Retain model-download failure causes and propagate unrelated exporter import failures.
• Fail immediately with the original cause when onnxruntime cannot be imported.

These implementation fixes were already in repository source; this release makes them available in the standalone PyPI package. No recognition algorithm change is included.

Verification

• Installed wheel and source distribution passed the actual wrapper regression suite on Python 3.11 and 3.12 in exact-head CI.
• Public PyPI wheel/sdist SHA256 match the accepted candidates; post-publication installation and error-contract tests passed on Python 3.12.
• ONNX inference remains torch-free. These checks are not an acoustic-accuracy, GPU or performance benchmark.

Both wheel and source archive are attached, with SHA256SUMS and BUILD-INFO.json. Source: signed tag at commit 5145ba3; PR #3690. Issue #3523 remains open for original-environment verification.

PyPI package | Main FunASR toolkit releases

FunASR v1.4.14

Highlights

• Fix portable source archives for Git archive and source installs.
• Keep MOSS-Transcribe-Diarize discoverable across models, deployment, and service docs.
• Ship the current realtime and industrial deployment guidance.

Install: python -m pip install -U "funasr==1.4.14"

This release includes verified runtime-llamacpp-v0.2.6 cross-platform binaries alongside the wheel and source package. SHA256SUMS covers every attached asset.

Full changelog: v1.4.13...v1.4.14

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==1.4.14"

v1.4.13

What's Changed

• docs: align localized deployment release references by @LauraGPT in #3614
• fix(onnx): preserve missing runtime dependency errors by @LauraGPT in #3605
• fix(fun-asr-nano): normalize ISO language hints by @LauraGPT in #3613
• feat(server): support n8n OpenAI audio transcription by @LauraGPT in #3612
• fix(realtime): accept bounded partial VAD overrun by @LauraGPT in #3591
• fix: prevent NumPy 2 ABI mismatches by @LauraGPT in #3615
• docs: clarify native vLLM transcription path by @LauraGPT in #3603
• docs: surface MOSS in use-case selection by @LauraGPT in #3608
• docs: clarify Fun-ASR-Nano HF CTC boundary by @LauraGPT in #3609
• docs: add Open WebUI MOSS service recipe by @LauraGPT in #3607
• docs(vllm): add measured L20 tuning profile by @LauraGPT in #3598
• docs: clarify MOSS speaker-label boundaries by @LauraGPT in #3590
• docs: surface current runtime packages in ja and ko by @LauraGPT in #3604
• feat(qwen3-asr): add offline vLLM example by @LauraGPT in #3592
• docs: make the roadmap actionable by @LauraGPT in #3593
• docs: refresh ecosystem growth snapshot by @LauraGPT in #3606

Full Changelog: v1.4.12...v1.4.13

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==1.4.13"

v1.4.12

What's Changed

• fix(realtime): preserve aligned long-segment partials by @LauraGPT in #3589
• docs: keep README updates focused by @LauraGPT in #3594
• feat: serve MOSS diarization across deployment stack by @LauraGPT in #3599
• fix(fun-asr-nano): stabilize fp16 vLLM decoding by @LauraGPT in #3597
• release: prepare FunASR 1.4.12 by @LauraGPT in #3602

Full Changelog: v1.4.11...v1.4.12

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==1.4.12"

Verify downloads

All release assets are listed in SHA256SUMS. Run sha256sum -c SHA256SUMS after downloading the files you need.

v1.4.11

What's Changed

• ci(release): attach Python packages automatically by @LauraGPT in #3585
• fix(auto-model): align SentencePiece word markers by @LauraGPT in #3587
• release: prepare FunASR 1.4.11 by @LauraGPT in #3588

Full Changelog: v1.4.10...v1.4.11

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==1.4.11"

FunASR 1.4.10

FunASR 1.4.10

This release fixes readable subtitle fallback boundaries for overlong Chinese sentences. Instead of mechanically filling each cue to the 42-character limit, FunASR now balances duration and length while preferring Jieba lexical boundaries, punctuation, script transitions, and real timestamp gaps.

On the reporter exact 10m33s mixed Chinese/Korean sample, 84 source segments became 133 cues with:

• maximum duration: 7,860 ms
• maximum length: 42 characters
• zero duration or length violations
• phrases such as 钟书成为 and 扭了一下 preserved

Issue #3539 remains open pending reporter confirmation on the published package.

Install

python -m pip install -U "funasr==1.4.10"

Python package SHA-256

b520f5f91e1f7ae86fd6f8fffb0980f7a9753ad34d2a13d8a9846caa60bcda19 funasr-1.4.10-py3-none-any.whl
67dd08ba621459157de7609ed481d5341bb8b26c23b1f745cc31bef33cf61ae4 funasr-1.4.10.tar.gz

The release also includes the verified llama.cpp runtime archives for Linux, macOS, and Windows, including CPU, Vulkan, CUDA, and Blackwell variants where supported.

Fix: #3583
Build workflow: https://github.com/modelscope/FunASR/actions/runs/33333569782

v1.4.9

What's Changed

• fix(subtitles): split overlong source cues by @LauraGPT in #3574
• fix(auto-model): preserve explicit timestamp and speaker text by @John6666cat in #3553
• fix(paraformer): trim timestamp predictor padding per sample by @John6666cat in #3554
• fix(realtime): preserve long segment tails by @LauraGPT in #3576
• docs: add native SGLang path for MOSS diarization by @LauraGPT in #3577
• chore(release): prepare 1.4.9 by @LauraGPT in #3578

New Contributors

• @John6666cat made their first contribution in #3553

Full Changelog: v1.4.8...v1.4.9

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==1.4.9"