diff --git a/README.md b/README.md index 0681766e..6482d861 100644 --- a/README.md +++ b/README.md @@ -1,3 +1,42 @@ +## 天数智芯 天垓100 sherpa-onnx + +### 编译指令 +在corex:3.2.1镜像中使用以下命令 +``` +SHERPA_ONNX_CMAKE_ARGS='-DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DSHERPA_ONNX_ENABLE_GPU=ON' python3 setup.py bdist_wheel +``` +即可编译出适用于天垓100的sherpa-onnx whl包 + +### pip使用方式 +```shell +pip install sherpa_onnx-1.12.5+corex3.2.1-cp310-cp310-linux_x86_64.whl +``` +具体的推理过程代码与sherpa-onnx python包没有任何区别 + +### 天垓100的模型适配情况 +我们在天垓100上针对sherpa-onnx offline asr部分进行了所有大类的适配,测试方式为在CPU和天垓100加速卡上对同一段长音频进行语音识别任务,获取运行时间,1-cer指标。运行环境为: +- CPU: Intel(R) Xeon(R) Gold 6238R CPU @ 2.20GHz 使用4线程推理 +- 天垓100:在上述CPU硬件的同时下使用一张天垓100加速卡 + +| 模型大类 | 模型名称 |CPU运行时间(秒)|天垓100运行时间(秒)|CPU 1-cer|天垓100 1-cer| 备注 | +|------|---------------|-----|----|-------|-------|---------------------| +| sense_voice |sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17| 36.8381 | 3.1053 | 0.990017 | 0.990017 | | +| whisper | sherpa-onnx-whisper-large-v3 | 705.8439 | 204.0064 | 0.722130 | 0.732113 | 天垓100的识别准确率比CPU略高 | +| whisper | sherpa-onnx-whisper-base | 34.8192 | 22.2912 | 0.682196 | 0.682196 | | +| whisper | sherpa-onnx-whisper-small | 119.6921 | 46.7079 | 0.662230 | 0.662230 | | +| whisper | sherpa-onnx-whisper-tiny | 18.2531 | 15.2332 | 0.584027 | 0.584027 | | +| whisper | sherpa-onnx-whisper-medium-aishell | 418.8124 | 141.6660 | 0.785358 | 0.783694 | | +| dolphin_ctc | sherpa-onnx-dolphin-small-ctc-multi-lang-2025-04-02 | 17.6391 | 2.7845 | 0.923461 | 0.923461 | | +| fire_red_asr | sherpa-onnx-fire-red-asr-large-zh_en-2025-02-16 | 409.8351 | 120.9066 | 0.960067 | 0.960067 | | +| zipformer_ctc | sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03 | 10.5072 | 6.9715 | 0.963394 | 0.966722 | 天垓100的识别准确率比CPU略高 | +| zipformer_ctc | sherpa-onnx-zipformer-ctc-small-zh-2025-07-16 | 4.9449 | 1.7400 | 0.958403 | 0.908486 | 天垓100在该模型与CPU相比精度有一些欠缺 | +| paraformer | sherpa-onnx-paraformer-zh-2024-03-09 | 6.4983 | 15.9283 | 0.948419 | 0.948419 | 天垓100运行速度明显慢于4线程CPU推理 | +| telespeech_ctc | sherpa-onnx-telespeech-ctc-zh-2024-06-04 | 24.0849 | 72.0935 | 0.948419 | 0.948419 | 天垓100运行速度明显慢于4线程CPU推理 | +| wenet_ctc | sherpa-onnx-zh-wenet-aishell2 | 3.9872 | 3.7492 | 0.813644 | 0.813644 | | +| transducer | sherpa-onnx-zipformer-multi-zh-hans-2023-9-2 | 6.4223 | 2.6784 | 0.920133 | 0.765391 | 天垓100在该模型与CPU相比精度有较大损失 | +| nemo_ctc | sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000 | N/A | N/A | N/A | N/A | 无法适配 | +| nemo_canary | sherpa-onnx-nemo-canary-180m-flash-en-es-de-fr | N/A | N/A | N/A | N/A | 无法适配 | +| moonshine | sherpa-onnx-moonshine-base-en-int8 | N/A | N/A | N/A | N/A | 无法适配 | ### Supported functions |Speech recognition| [Speech synthesis][tts-url] | [Source separation][ss-url] |