Recognition of bimodal produced speech based on Support Vector Machines
Jovan Galić, Dragana Šumarac Pavlović, Slobodan T. Jovičić, Branko Marković, Đorđe Grozdić · 2017
This paper presents the results of experiments in recognition of whispered speech, as a specific mode of verbal communication, using SVM (Support Vector Machines) classifier. For training and testing purposes, part of the speech database Whi-Spe with male speakers has been exploited. In matched scenarios average word recognition rate was 99.3% (for normally phonated speech) and 97.8% (for whisper). In mismatched scenarios, recognition of whisper with training on normal speech utterances was 75.4%, whereas recognition of whisper-trained normal speech was with recognition rate of 81.3%.