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

```bash
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](https://github.com/modelscope/FunASR/blob/v1.4.15/examples/industrial_data_pretraining/qwen3_asr/transcribe_vllm_offline.py) is a source recipe, not a newly installed wheel command. Follow the [versioned MOSS guide](https://github.com/modelscope/FunASR/blob/v1.4.15/docs/moss_transcribe_diarize.md) 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](https://github.com/modelscope/FunASR/releases/tag/v1.4.14). Downgrading FunASR alone does not restore other changed dependencies.

[Full source changes](https://github.com/modelscope/FunASR/compare/v1.4.14...v1.4.15)

<!-- funasr-runtime-downloads:start -->
## Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: [runtime-llamacpp-v0.2.6](https://github.com/modelscope/FunASR/releases/tag/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](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-linux-arm64.tar.gz) | `7bca29cfa3c9a08e235a62212ca9e00f6656e59a8f07078966a2bfda1e5aa1f9` |
| Linux x64 AVX2 | [funasr-llamacpp-linux-x64-avx2.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-linux-x64-avx2.tar.gz) | `aaebc5470f846ce915200b35d6e9f9bd0a0d3ed399d39e49bdeb7a1f1782bc70` |
| Linux x64 Vulkan | [funasr-llamacpp-linux-x64-vulkan.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-linux-x64-vulkan.tar.gz) | `f02d41e98e9d4041f0896661007193810f025484d2175958f7c1313d5c90ec46` |
| Linux x64 portable | [funasr-llamacpp-linux-x64.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-linux-x64.tar.gz) | `779967de1c528c2be966bcc47f246e7d3e6fcdb748d9491263062f4120f35e52` |
| macOS arm64 | [funasr-llamacpp-macos-arm64.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-macos-arm64.tar.gz) | `bda59474202b887190f59d25b7b42c714469efae71276072c12fa0a38de68792` |
| Windows x64 AVX2 | [funasr-llamacpp-windows-x64-avx2.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-windows-x64-avx2.zip) | `062cda8fefadd31c3e811227116daccf448a8520f4b0bb168d225c896e65ebbd` |
| Windows x64 CUDA Blackwell (sm_120) | [funasr-llamacpp-windows-x64-cuda-blackwell.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-windows-x64-cuda-blackwell.zip) | `e32961a753f40888182f352fa551159c5165a6a77718ae4ade316aedfea4b1c2` |
| Windows x64 CUDA | [funasr-llamacpp-windows-x64-cuda.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-windows-x64-cuda.zip) | `148657911fb666b7af6ec43af2e23a0984e3259012b4c39f95631b717feb6840` |
| Windows x64 Vulkan | [funasr-llamacpp-windows-x64-vulkan.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.15/funasr-llamacpp-windows-x64-vulkan.zip) | `debf8007e55011cad06081e7b8a78972f1b8fe672bc324d41e650d68821f6a6a` |
| Windows x64 portable | [funasr-llamacpp-windows-x64.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.15/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:

```bash
python -m pip install -U "funasr==1.4.15"
```
<!-- funasr-runtime-downloads:end -->

## Artifact verification

[SHA256SUMS](https://github.com/modelscope/FunASR/releases/download/v1.4.15/SHA256SUMS) covers the wheel, sdist and ten native archives. [PROVENANCE.json](https://github.com/modelscope/FunASR/releases/download/v1.4.15/PROVENANCE.json) records their independent sources and verification scope. PyPI and GitHub Python artifact bytes match.