热词间有相互干扰 - Githubissues

modelscope / FunASR

A Fundamental End-to-End Speech Recognition Toolkit and Open Source SOTA Pretrained Models, Supporting Speech Recognition, Voice Activity Detection, Text Post-processing etc.

https://www.funasr.com

Other

7.19k stars 764 forks source link

热词间有相互干扰 #1727

Closed kli017 closed 6 months ago

kli017 commented 6 months ago

Notice: In order to resolve issues more efficiently, please raise issue following the template. （注意：为了更加高效率解决您遇到的问题，请按照模板提问，补充细节）

🐛 Bug

在runtime环境下使用speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404-onnx模型。添加如下热词表时感觉热词直接会有相互干扰。比如针灸铜人 80 久通 80

测试可能会出现针灸通人、久铜等结果。请问添加热词是单独提高token概率吗。如果是全词匹配的话，按说wfst里影响不会这么大。有无办法解决？

Environment

OS (e.g., Linux): Linux
FunASR Version (e.g., 1.0.0): FunASR
ModelScope Version (e.g., 1.11.0): runtime
PyTorch Version (e.g., 2.0.0):
How you installed funasr (pip, source):
Python version:
Docker version: funasr-runtime-sdk-online-cpu-0.1.9
Any other relevant information:

R1ckShi commented 6 months ago

runtime中的热词分两部分，首先是基于clas的nn热词，这个阶段是通过attention进行热词与decoder信息的匹配的有热词冲突会导致attention机制产生错误的相关性，没有很好的解法可能的解决方法是拆解长热词或者把短热词补长