v1.3.10

FunASR v1.3.10

New features

• Agent-friendly CLI: funasr audio.wav --output-format json for structured output
• Fun-ASR-Nano: batched VAD-segment decoding (~1.75× faster) (#2979)
• WebSocket 2-pass server: sentence-level timestamps
• serve_vllm.py: new --vad-model / --spk-model flags

Fixes

• Fun-ASR-Nano: bf16/fp16 inference no longer crashes; warn on degraded fp16 (#2980)
• Fun-ASR-Nano vLLM: fix CUDA crash from repetition_penalty
• CLI: valid SRT timestamps + correct JSON durations (#2982); use sentence_info text (#2983); correct model id Fun-ASR-Nano-2512 (#2984)
• Clearer error for missing audio path (#2981); respect explicit VAD silence threshold; handle None encoder/scheduler configs

Docs

• New CLI reference; clearer vLLM install guidance

Full changelog: v1.3.9...v1.3.10

v1.3.9: Wheel packaging + SenseVoice speaker diarization fix

What's New

Wheel packaging (fixes #2943)

FunASR now publishes a py3-none-any wheel alongside the source distribution. Installation is faster since pip no longer needs to build from source.

Bug fixes

• SenseVoice + speaker diarization: Fixed crash when using spk_model="cam++" with SenseVoice (auto-falls back to VAD-segment mode since SenseVoice doesn't produce word-level timestamps)
• torchaudio >= 2.11 compatibility: Added soundfile as intermediate fallback for users with newer torchaudio versions that removed legacy backends

Install / Upgrade

pip install --upgrade funasr

Full changelog: v1.3.3...v1.3.9

v1.3.3: Agent Integration — OpenAI API + MCP Server + funasr-server CLI

Highlights

This release makes FunASR a drop-in speech backend for AI agents.

New: funasr-server CLI

pip install funasr fastapi uvicorn python-multipart
funasr-server --device cuda

One command starts an OpenAI-compatible /v1/audio/transcriptions endpoint.

New: MCP Server

AI assistants (Claude, Cursor, Windsurf) can now transcribe audio directly.

New: OpenAI-Compatible API

Works with any agent framework: LangChain, AutoGen, CrewAI, Dify, Flowise, Open WebUI.

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")
result = client.audio.transcriptions.create(model="sensevoice", file=open("a.wav","rb"))

Bug Fixes

• Fixed hub="hf" parameter propagation to sub-models (v1.3.2)
• Fixed Qwen3-ASR ImportError masking

Upgrade

pip install --upgrade funasr

Links

• Agent Integration Guide
• OpenAI API Docs
• MCP Server Docs
• Benchmark

v1.3.2: HuggingFace Hub Fix + Performance Benchmark

What's New

Bug Fix

• Fixed hub parameter propagation — When using hub="hf", the parameter is now correctly forwarded to VAD/PUNC/SPK sub-models. Previously, users on HuggingFace would get 404 errors for sub-models. (#2859)

Improvements

• Updated PyPI metadata with better description, keywords, and project URLs
• Added comprehensive benchmark page: https://modelscope.github.io/FunASR/benchmark.html

Benchmark Results (PyTorch, GPU)

Model	Type	Speed
SenseVoice-Small	NAR	170x realtime
Paraformer-Large	NAR	120x realtime
Whisper-large-v3-turbo	AR	46x realtime
Fun-ASR-Nano	LLM	17x realtime
Whisper-large-v3	AR	13.4x realtime

Install / Upgrade

pip install --upgrade funasr

Quick Start

from funasr import AutoModel
model = AutoModel(model="FunAudioLLM/SenseVoiceSmall", hub="hf", vad_model="funasr/fsmn-vad", device="cuda")
result = model.generate(input="audio.wav")

0.3.0

What's new:

2023.3.17, funasr-0.3.0, modelscope-1.4.1

• New Features:
  • Added support for GPU runtime solution, nv-triton, which allows easy export of Paraformer models from ModelScope and deployment as services. We conducted benchmark tests on a single GPU-V100, and achieved an RTF of 0.0032 and a speedup of 300.
  • Added support for CPU runtime quantization solution, which supports export of quantized ONNX and Libtorch models from ModelScope. We conducted benchmark tests on a CPU-8369B, and found that RTF increased by 50% (0.00438 -> 0.00226) and double speedup (228 -> 442).
  • Added support for C++ version of the gRPC service deployment solution. The C++ version of ONNXRuntime and quantization solution, provides double higher efficiency compared to the Python runtime, demo.
  • Added streaming inference pipeline to the 16k VAD model, 8k VAD model, with support for audio input streams (>= 10ms) , demo.
  • Improved the punctuation prediction model, resulting in increased accuracy (F-score increased from 55.6 to 56.5).
  • Added real-time subtitle example based on gRPC service, using a 2-pass recognition model. Paraformer streaming model is used to output text in real time, while Paraformer-large offline model is used to correct recognition results, demo.
• New Models:
  • Added 16k Paraformer streaming model, which supports real-time speech recognition with streaming audio input, demo. It can be deployed using the gRPC service to implement real-time subtitle function.
  • Added streaming punctuation model, which supports real-time punctuation marking in streaming speech recognition scenarios, with real-time calls based on VAD points. It can be used along with real-time ASR models to achieve readable real-time subtitle function, demo.
  • Added TP-Aligner timestamp model, which takes audio and corresponding text as input and outputs word-level timestamps. Its performance is comparable to that of the Kaldi FA model (60.3ms vs. 69.3ms). It can be combined freely with ASR models, demo.
  • Added financial domain model (8k Paraformer-large-3445vocab), which is fine-tuned using 1000 hours of data. The recognition accuracy on the financial domain test set increased by 5%, and the recall rate of domain keywords increased by 7%.
  • Added audio-visual domain model (16k Paraformer-large-3445vocab), which is fine-tuned using 10,000 hours of data. The recognition accuracy on the audio-visual domain test set increased by 8%.
  • Added 8k speaker verification model, which can be used for speaker embedding extraction.
  • Added speaker diarization models, including 16k SOND Chinese model, 8k SOND English model, which achieved the best performance on AliMeeting and Callhome with a DER of 4.46% and 11.13%, respectively.
  • Added UniASR streaming offline unifying models, including 16k UniASR Burmese, 16k UniASR Hebrew, 16k UniASR Urdu, 8k UniASR Mandarin financial domain, and 16k UniASR Mandarin audio-visual domain.

最新更新：

• 2023年3月17日：funasr-0.3.0, modelscope-1.4.1
  • 功能完善：
    • 新增GPU runtime方案，nv-triton，可以将modelscope中Paraformer模型便捷导出，并部署成triton服务，实测，单GPU-V100，RTF为0.0032，吞吐率为300，benchmark。
    • 新增CPU runtime量化方案，支持从modelscope导出量化版本onnx与libtorch，实测，CPU-8369B，量化后，RTF提升50%（0.00438->0.00226），吞吐率翻倍（228->442），benchmark。
    • 新增加C++版本grpc服务部署方案，配合C++版本onnxruntime，以及量化方案，相比python-runtime性能翻倍。
    • 16k VAD模型，8k VAD模型，modelscope pipeline，新增加流式推理方式，，最小支持10ms语音输入流，用法。
    • 优化标点预测模型，主观体验标点准确性提升(fscore绝对提升 55.6->56.5)。
    • 基于grpc服务，新增实时字幕demo，采用2pass识别模型，Paraformer流式模型 用来上屏，Paraformer-large离线模型用来纠正识别结果。
  • 上线新模型：
    • 16k Paraformer流式模型，支持语音流输入，可以进行实时语音识别，用法。支持基于grpc服务进行部署，可实现实时字幕功能。
    • 流式标点模型，支持流式语音识别场景中的标点打标，以VAD点为实时调用点进行流式调用。可与实时ASR模型配合使用，实现具有可读性的实时字幕功能，用法
    • TP-Aligner时间戳模型，输入音频及对应文本输出字级别时间戳，效果与Kaldi FA模型相当（60.3ms v.s. 69.3ms），支持与asr模型自由组合，用法。
    • 金融领域模型，8k Paraformer-large-3445vocab，使用1000小时数据微调训练，金融领域测试集识别效果相对提升5%，领域关键词召回相对提升7%。
    • 音视频领域模型，16k Paraformer-large-3445vocab，使用10000小时数据微调训练，音视频领域测试集识别效果相对提升8%。
    • 8k说话人确认模型，CallHome数据集英文说话人确认模型，也可用于声纹特征提取。
    • 说话人日志模型，16k SOND中文模型，8k SOND英文模型，在AliMeeting和Callhome上获得最优性能，DER分别为4.46%和11.13%。
    • UniASR流式离线一体化模型:
      16k UniASR缅甸语、 16k UniASR希伯来语、 16k UniASR乌尔都语、 8k UniASR中文金融领域、16k UniASR中文音视频领域。

New Contributors

• @dingbig made their first contribution in #147
• @yuekaizhang made their first contribution in #161
• @zhuz...

v0.2.0

What's new:

2023.2.17, funasr-0.2.0, modelscope-1.3.0

• We support a new feature, export paraformer models into onnx and torchscripts from modelscope. The local finetuned models are also supported.
• We support a new feature, onnxruntime, you could deploy the runtime without modelscope or funasr, for the paraformer-large model, the rtf of onnxruntime is 3x speedup(0.110->0.038) on cpu, details.
• We support a new feature, grpc, you could build the ASR service with grpc, by deploying the modelscope pipeline or onnxruntime.
• We release a new model paraformer-large-contextual, which supports the hotword customization based on the incentive enhancement, and improves the recall and precision of hotwords.
• We optimize the timestamp alignment of Paraformer-large-long, the prediction accuracy of timestamp is much improved, and achieving accumulated average shift (aas) of 74.7ms, details.
• We release a new model, 8k VAD model, which could predict the duration of none-silence speech. It could be freely integrated with any ASR models in modelscope.
• We release a new model, MFCCA, a multi-channel multi-speaker model which is independent of the number and geometry of microphones and supports Mandarin meeting transcription.
• We release several new UniASR model: Southern Fujian Dialect model, French model, German model, Vietnamese model, Persian model.
• We release a new model, paraformer-data2vec model, an unsupervised pretraining model on AISHELL-2, which is inited for paraformer model and then finetune on AISHEL-1.
• We release a new feature, the VAD, ASR and PUNC models could be integrated freely, which could be models from modelscope, or the local finetine models. The demo.
• We optimize punctuation common model, enhance the recall and precision, fix the badcases of missing punctuation marks.
• Various new types of audio input types are now supported by modelscope inference pipeline, including: mp3、flac、ogg、opus...

最新更新：

• 2023年2月（2月17号发布）：funasr-0.2.0, modelscope-1.3.0

  • 功能完善：

    • 新增加模型导出功能，Modelscope中所有Paraformer模型与本地finetune模型，支持一键导出onnx格式模型与torchscripts格式模型，用于模型部署。
    • 新增加Paraformer模型onnxruntime部署功能，无须安装Modelscope与FunASR，即可部署，cpu实测，onnxruntime推理速度提升近3倍(rtf: 0.110->0.038)。
    • 新增加grpc服务功能，支持对Modelscope推理pipeline进行服务部署，也支持对onnxruntime进行服务部署。
    • 优化Paraformer-large长音频模型时间戳，对badcase时间戳预测准确率有较大幅度提升，平均首尾时间戳偏移74.7ms，详见论文。
    • 新增加任意VAD模型、ASR模型与标点模型自由组合功能，可以自由组合Modelscope中任意模型以及本地finetune后的模型进行推理，用法示例。
    • 优化标点通用模型，增加标点召回和精度，修复缺少标点等问题。
    • 新增加采样率自适应功能，任意输入采样率音频会自动匹配到模型采样率；新增加多种语音格式支持，如，mp3、flac、ogg、opus等。

  • 上线新模型：

    • Paraformer-large热词模型，可实现热词定制化，基于提供的热词列表，对热词进行激励增强，提升模型对热词的召回。
    • MFCCA多通道多说话人识别模型，与西工大音频语音与语言处理研究组合作论文，一种基于多帧跨通道注意力机制的多通道语音识别模型。
    • 8k语音端点检测VAD模型，可用于检测长语音片段中有效语音的起止时间点，支持流式输入，最小支持10ms语音输入流。
    • UniASR流式离线一体化模型: 16k UniASR闽南语、 16k UniASR法语、 16k UniASR德语、 16k UniASR越南语、 16k UniASR波斯语。
    • 基于Data2vec结构无监督预训练Paraformer模型，采用Data2vec无监督预训练初值模型，在AISHELL-1数据中finetune Paraformer模型。

New Contributors

• @zjc6666 made their first contribution in #35
• @lyblsgo made their first contribution in #37
• @lingyunfly made their first contribution in #42
• @fangd123 made their first contribution in #44
• @dyyzhmm made their first contribution in #48
• @R1ckShi made their first contribution in #50
• @chenmengzheAAA made their first contribution in #57
• @ZhihaoDU made their first contribution in #95
• @SWHL made their first contribution in #97
• @yufan-aslp made their first contribution in #105
• @magicharry made their first contribution in #119

Full Changelog: v0.1.6...v0.2.0

v0.1.6

Release Notes:

2023.1.16, funasr-0.1.6

• We release a new version model Paraformer-large-long, which integrate the VAD model, ASR,
  Punctuation model and timestamp together. The model could take in several hours long inputs.
• We release a new type model, VAD, which could predict the duration of none-silence speech. It could be freely integrated with any ASR models in Model Zoo.
• We release a new type model, Punctuation, which could predict the punctuation of ASR models's results. It could be freely integrated with any ASR models in Model Zoo.
• We release a new model, Data2vec, an unsupervised pretraining model which could be finetuned on ASR and other downstream tasks.
• We release a new model, Paraformer-Tiny, a lightweight Paraformer model which supports Mandarin command words recognition.
• We release a new type model, SV, which could extract speaker embeddings and further perform speaker verification on paired utterances. It will be supported for speaker diarization in the future version.
• We improve the pipeline of modelscope to speedup the inference, by integrating the process of build model into build pipeline.
• Various new types of audio input types are now supported by modelscope inference pipeline, including wav.scp, wav format, audio bytes, wave samples...

最新更新

• 2023年1月（1月16号发布）：funasr-0.1.6, modelscope-1.2.0
  • 上线新模型：
    • Paraformer-large长音频模型，集成VAD、ASR、标点与时间戳功能，可直接对时长为数小时音频进行识别，并输出带标点文字与时间戳。
    • 中文无监督预训练Data2vec模型，采用Data2vec结构，基于AISHELL-2数据的中文无监督预训练模型，支持ASR或者下游任务微调模型。
    • 16k语音端点检测VAD模型，可用于检测长语音片段中有效语音的起止时间点。
    • 中文标点预测通用模型，可用于语音识别模型输出文本的标点预测。
    • 8K UniASR流式模型，8K UniASR模型，一种流式与离线一体化语音识别模型，进行流式语音识别的同时，能够以较低延时输出离线识别结果来纠正预测文本。
    • Paraformer-large基于AISHELL-1微调模型、AISHELL-2微调模型，将Paraformer-large模型分别基于AISHELL-1与AISHELL-2数据微调。
    • 说话人确认模型 ，可用于说话人确认，也可以用来做说话人特征提取。
    • 小尺寸设备端Paraformer指令词模型，Paraformer-tiny指令词版本，使用小参数量模型支持指令词识别。
  • 将原TensorFlow模型升级为Pytorch模型，进行推理，并支持微调定制，包括：
    • 16K 模型：Paraformer中文、Paraformer-large中文、UniASR中文、UniASR-large中文、UniASR中文流式模型、UniASR方言、UniASR方言流式模型、UniASR日语、UniASR日语流式模型、UniASR印尼语、UniASR印尼语流式模型、UniASR葡萄牙语、UniASR葡萄牙语流式模型、UniASR英文、UniASR英文流式模型、UniASR俄语、UniASR俄语流式模型、UniASR韩语、UniASR韩语流式模型、UniASR西班牙语、UniASR西班牙语流式模型、UniASR粤语简体、UniASR粤语简体流式模型、
    • 8K 模型：Paraformer中文、UniASR中文、UniASR中文流式模型

New Contributors

• @nichongjia-2007 made their first contribution in #27

Full Changelog: v0.1.4...v0.1.6

v0.1.4

The is the first release version.

1. Paraformer model could be decoding with batch >1.
2. UniASR model and recipes are new added.
3. Transformer and Conformer are also contained.
4. The inference and finetuning of models in modelscope are more convenience.