search

NickRuiz / power-asr

Phonetically-Oriented Word Error Rate
MIT License
33 stars 9 forks source link

algorithm running infinitely #6

Open romulomello opened 5 months ago

romulomello commented 5 months ago

When there is a reference a little far from the hypothesis, it seems that the code runs infinitely. Do you know of any place I can try to adjust to improve this issue a little? I use your code to align outputs from an asr model with references and these texts are not always similar.

Another question is that in the readme you mention "Alternative implementation of phoneme alignments with CMUDict (http://www.speech.cs.cmu.edu/cgi-bin/cmudict) and Hunspell for syllabification (e.g. https://pyphen. org/) instead of Festival", how do I run this alternative version? Because my model's output is in phonemes "IPA", I think it would do better then.

romulomello commented 5 months ago

I forgot to comment but I'm using it to align texts in Portuguese - Brazilian