ASR English model cannot recognize my voice well

[jims57]

why all the english related model in asr cannot recognize well? though Chinese ones seem work well.

Have you tested the english related models? Am i missing anything to make it accurate? thanks.

e.g. this model:sherpa-ncnn-streaming-zipformer-20m-2023-02-17

[jims57]

I don't think it is about my english tone, because even i use the standard english audio, it doesn't work well. the recognized result seem bad.

[csukuangfj]

please list all.models.you have tested

You only list one.model.

[jims57]

i have test all the models listed sherpa, i don't think any of them is good enough to recognize english, though chinese one is good

[jims57]

btw, please see : https://k2-fsa.github.io/icefall/model-export/export-ncnn-conv-emformer.html#export-the-model-via-torch-jit-trace

[jims57]

i have test the model in both onnx and ncnn examples for swift, they cannot recognize English well, i can say it is very bad, WER is at least 50%, am i missing anything? can you try the model using the swift sample for a try? thanks.

[csukuangfj]

【新一代 Kaldi: 最新版 zipformer 在 iOS 上的英文语音识别演示-哔哩哔哩】 https://b23.tv/K87JkfW

【新一代 Kaldi: Two-pass 实时英语语音识别之 iOS （奥巴马演讲)-哔哩哔哩】 https://b23.tv/YS2QNmR

We have already provided Engine asr video demos

If it is not working for you, please recheck.