Application of Neural Networks in Whispered Speech Recognition

Abstract

This paper presents the preliminary results of experimental research of whispered speech recognition that was based on the application of artificial neural networks (ANN). The paper also describes a speech database of words that were spoken in a whisper and normal manner, which was created especially for this study. A part of this database was used for preliminary training and testing of the ANN. Mel Frequency Cepstral Coefficients (MFCC) in normal speech and whispered speech were used as an input to the ANN. The case of speaker dependent recognition was tested and ANNs with optimal topologies have a 97.98% accuracy in speech recognition and 96.21% in whisper recognition for a male speaker. The results for a female speaker were very similar. In the case of whisper recognition, when ANN was trained for normal speech the score of whisper recognition was 75.71% for a male speaker (82.14% for female), and vice versa, when ANN was trained for whisper, normal speech recognition was 82.14% for a male speaker (90% for female).

Keywords

• Neural networks
• speech recognition
• whisper recognition
• whispered speech database