Binary Neural Networks for Classification of Voice Commands From Throat Microphone

Abstract

Multi-class pattern classification has many applications including speech recognition, and it is not easy to extend from two-class neural networks (NNs). This paper presents a study about using binary classifiers with NNs together with a perceptual linear prediction (PLP) method for feature extraction to increase the classification rate of voice commands captured using a throat microphone, comparing this method with a single NN. Because there is no other data set with voice commands captured using a throat microphone in the Brazilian Portuguese language in researched literature, we created a data set with isolated voice commands with utterances captured from 150 people (men and women). All the voice samples are captured in Brazilian Portuguese, and they are the digits “0”through “9”and the words “Ok”and “Cancel”. The results show that the throat microphone is robust in noise environment, achieving 95.4% of hit rate in our speech recognition system with multiple NNs using the one-against-all approach, better performance than a simple NN that reach 91.88%. This result is very representative, since both classifiers obtained high hit rates. But, it requires 535% more time for training the multiple NNs compared with simple NN. The best configuration on PLP extraction order is 9 or 10 for voice samples captured by the throat microphone, which was observed that poor stressed vowel and fricative-like words “3” and “7”in Portuguese confuses the classifier.

Keywords

• Multi-class pattern recognition
• speech recognition
• neural networks
• binary classifiers