--- title: alison-lingsheng-asr canonical_url: "https://www.modelscope.cn/models/xixiaxc/alison-lingsheng-asr" md_url: "https://www.modelscope.cn/models/xixiaxc/alison-lingsheng-asr.md" repository: xixiaxc/alison-lingsheng-asr chinese_name: "Alison Workspace 中文精准识别模型" last_updated: 2026-09-24 license: "Apache License 2.0" pipeline_tag: auto-speech-recognition tasks: - auto-speech-recognition library_name: - onnx frameworks: - onnx language: - zh - yue downloads: 0 stars: 0 tags: - Paraformer - SenseVoice - FunASR - CAM++ --- # alison-lingsheng-asr > alison-lingsheng-asr - xixiaxc 在 ModelScope 开源的模型。Alison Workspace 中文精准识别模型包（v1） xixiaxc/alison-lingsheng-asr 是 ModelScope 魔搭社区上的auto-speech-recognition模型，采用 Apache License 2.0 许可。 - **Repository**: xixiaxc/alison-lingsheng-asr - **License**: Apache License 2.0 - **Tasks**: auto-speech-recognition - **Tags**: Paraformer, SenseVoice, FunASR, CAM++ - **Downloads**: 0 - **Stars**: 0 - **Last updated**: 2026-09-24 Source: https://www.modelscope.cn/models/xixiaxc/alison-lingsheng-asr --- # Alison Workspace 中文精准识别模型包（v1） > 同一套文件也在 Hugging Face：[wangxp9527/alison-lingsheng-asr](https://huggingface.co/wangxp9527/alison-lingsheng-asr)（App 优先从魔搭下载，失败时改从 Hugging Face 下载）。 macOS 应用 **Alison Workspace** 的「中文精准识别」功能按需下载的模型文件。App 内置的推理引擎（ONNX Runtime 1.19.2，CPU）直接读取本仓库 `v1/` 目录下的文件；它们不是为单独使用而整理的。 ## 两种模式 | 模式 | 实时草稿 | 整句定稿 | 热词 | 用到的目录 | |---|---|---|---|---| | 普通话 | 流式 Paraformer-large | 离线 Paraformer-large（Contextual） | 支持 | `v1/common` + `v1/mandarin` | | 方言增强 | 流式 Paraformer-large | SenseVoice-Small | 不支持 | `v1/common` + `v1/dialect` | 两种模式共用 FSMN-VAD（断句）、CT-Transformer（标点）、CAM++（说话人分离），都在 `v1/common`。 ## 出处与改动 所有模型的原作者是**阿里巴巴通义实验室**（FunASR、FunAudioLLM、3D-Speaker 团队），原始模型发布在魔搭社区。本仓库只做了格式转换与量化，没有重新训练，模型名称保持原样。 | 目录 | 原始模型 | 改动 | |---|---|---| | `common/online` | [Paraformer-large 流式](https://www.modelscope.cn/models/iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online-onnx) | `encoder.onnx` 即原仓库的 `model_quant.onnx`，未改；`decoder.onnx` 改了计算图里注意力掩码宽度的取法（改为取输入张量的长度维），以支持补零后的批量推理，权重未改，int8 | | `mandarin` | [Paraformer-large Contextual](https://www.modelscope.cn/models/iic/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404) | 拆成编码器、解码器两张图导出为 ONNX（同样修正掩码宽度），int8 量化；`hotword_eb.bin` 是热词功能用的权重，由该模型导出 | | `dialect` | [SenseVoice-Small](https://www.modelscope.cn/models/iic/SenseVoiceSmall) | 导出为 ONNX 后做 int8 动态量化（MatMul，逐通道）；`tokens.json`、`am.mvn` 与 [SenseVoiceSmall-onnx](https://www.modelscope.cn/models/iic/SenseVoiceSmall-onnx) 相同 | | `common/vad` | [FSMN-VAD](https://www.modelscope.cn/models/iic/speech_fsmn_vad_zh-cn-16k-common-onnx) | 权重转成 C++ 引擎直接读取的二进制格式（`vad_fsmn.bin`）；`am.mvn` 未改 | | `common/punc` | [CT-Transformer 中文标点](https://www.modelscope.cn/models/iic/punc_ct-transformer_zh-cn-common-vocab272727-onnx) | 词向量表中汉字与符号的行保持原精度，英文单词的行逐行量化为 int8（283 MB → 88 MB）；`tokens.json` 未改 | | `common/speaker` | [CAM++](https://www.modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common) | 导出为 ONNX（输入长度可变） | ## 许可 原始模型在魔搭社区的模型卡均标注 **Apache License 2.0**；SenseVoice-Small 在 Hugging Face 上另标注 [FunASR 模型开源协议 1.1](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)，该协议要求注明出处与作者、保留模型名称——本说明即为此声明。本仓库的文件按同样的许可分发。 ## 文件与校验 App 下载后逐个核对 SHA-256，不一致即丢弃重下。 | 文件 | 字节数 | SHA-256 | |---|---:|---| | `v1/common/online/am.mvn` | 11,203 | `29b3c740a2c0cfc6b308126d31d7f265fa2be74f3bb095cd2f143ea970896ae5` | | `v1/common/online/decoder.onnx` | 58,293,022 | `617de44512bded6d66025d2cc1b19aa92a03e45fe345194177b13be29dd332a3` | | `v1/common/online/encoder.onnx` | 166,350,528 | `dd4121cf45102018c26f9256f0b862df416edfcd06b0863ef4ce378a63c7d5e2` | | `v1/common/online/tokens.json` | 93,676 | `2b20c2b12572d682afff84ce1c8d560f67b8b32a4c1f21567411d141ed352127` | | `v1/common/punc/model.onnx` | 88,286,455 | `ae68ca42d87381bb4af015ec1c102fab7e8a974b7a659765fd01928dc92645b8` | | `v1/common/punc/tokens.json` | 4,207,480 | `c960ab87bccea4aa15cf49a59f71973c2c330b46668048cd8da253749ec71ee3` | | `v1/common/speaker/campplus.onnx` | 28,293,345 | `2680666824e4dfc428334d1fa9371bf1f61a491f920eabd89e846ba3c5642f50` | | `v1/common/vad/am.mvn` | 8,040 | `6820fef9687708c4fc3fab2530179c8fcea6262daa25514380056cd8f6eb1754` | | `v1/common/vad/vad_fsmn.bin` | 1,715,656 | `bdd7bd678507a2419653f852c3cb3bff6b6601cc7872a62f24244a7d6c9999a6` | | `v1/dialect/am.mvn` | 11,203 | `29b3c740a2c0cfc6b308126d31d7f265fa2be74f3bb095cd2f143ea970896ae5` | | `v1/dialect/model.onnx` | 241,217,824 | `c84c1b8b379b9ad89d48c70dacdc4d3cd9ef4c136b5b794279618459ffb6ce56` | | `v1/dialect/tokens.json` | 352,064 | `a2594fc1474e78973149cba8cd1f603ebed8c39c7decb470631f66e70ce58e97` | | `v1/mandarin/am.mvn` | 11,203 | `29b3c740a2c0cfc6b308126d31d7f265fa2be74f3bb095cd2f143ea970896ae5` | | `v1/mandarin/decoder.onnx` | 60,051,317 | `32eee3183840927d417b952d54d435d2a8e90aae9d037c7d1c739b754db6a425` | | `v1/mandarin/encoder.onnx` | 161,965,688 | `9e58531ceafbcefb635b9e853667dc937c1560c29b245fe5cea452ca10b3084d` | | `v1/mandarin/hotword_eb.bin` | 25,616,400 | `b2e34bd71308808363256bcfd33cf3c9d661c5bc1c9a4e596886934bd55ac778` | | `v1/mandarin/tokens.json` | 93,676 | `2b20c2b12572d682afff84ce1c8d560f67b8b32a4c1f21567411d141ed352127` | --- ## English On-demand model files for the **Precise Chinese Recognition** feature of the macOS app **Alison Workspace**. Two modes: *Mandarin* (streaming Paraformer-large drafts + Paraformer-large Contextual finals, hotwords supported) and *Dialect Enhanced* (streaming Paraformer-large drafts + SenseVoice-Small finals). Shared: FSMN-VAD, CT-Transformer punctuation, CAM++ speaker embeddings. All models are by **Alibaba Tongyi Lab** (FunASR / FunAudioLLM / 3D-Speaker) and were published on ModelScope under the Apache License 2.0 (SenseVoice-Small is additionally listed under the FunASR Model License 1.1 on Hugging Face). This repository only converts and quantizes them (ONNX export, int8 quantization, an attention-mask shape fix in the Paraformer decoder graphs, a C++ weight format for the VAD); no retraining. Original model names are kept. See the tables above for per-file origins, changes, sizes and SHA-256 checksums.