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.

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.

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.

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.

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

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.

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"

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.

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"

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.

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.

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.

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

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

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.

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"

What's Changed

• fix(download): preserve snapshot errors by @LauraGPT in #3565
• feat(moss): support vLLM diarized responses by @LauraGPT in #3566
• docs: refresh repository roadmap and issue boundaries by @LauraGPT in #3567
• feat(moss): support long vLLM transcriptions by @LauraGPT in #3568
• docs: expose MOSS LocalAI edge path by @LauraGPT in #3569
• feat(runtime): publish Windows CUDA asset for Blackwell by @LauraGPT in #3570
• docs(runtime): publish v0.2.6 download matrix by @LauraGPT in #3571
• docs(site): sync v0.2.6 and official MOSS recipe by @LauraGPT in #3572
• chore(release): prepare 1.4.8 by @LauraGPT in #3573

Full Changelog: v1.4.7...v1.4.8

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.

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==1.4.8"