Context biasing based on Aho–Corasick

[victor45664]

I fixed the bug described here. This PR can increase hotword recognition accuracy but it may also increase false positive. The optimal context_score should be around 3. But please adjust context_score according to your own develop set. And the RTF is also lower 0.3569-> 0.33.

[robin1001]

Any result about the accuracy?

[victor45664]

Any result about the accuracy?

wer 23.1 -> 20.01 on our private hotword dataset.

[robin1001]

any WER comparison about the AC and WFST?

[victor45664]

any WER comparison about the AC and WFST?

The WER without context biasing is 33 . WFST is 23.1 while AC is 20.01.

[robin1001]

how many context phrases in your testing?

[victor45664]

how many context phrases in your testing?

1216

[robin1001]

感谢支持，不好意思，这个 PR 在这里耽搁了太久了，这个是我们的内部实现，https://github.com/wenet-e2e/wenet/pull/1937，基于 FST 框架实现，和 wenet 更吻合一些，目前已合并。这个工作差不多也是和你这个同期做好的，但因为论文中要用，所以一直没有提上来，和你这个 PR 功能比较类似，所以没有持续跟进你这个 PR。