Detection of Fricatives in Continuous Speech Using Auto Resonance Networks

Fricatives are a class of noisy speech sounds that are produced by forming partial constriction in the vocal tract. As a result, they have dominant high-frequency contents. Difficulty in speech production could lead to the production of fricatives with a place of constriction different from the actual. Automatic detection of fricatives helps in identifying the fricative regions and modifying the speech segment in that region. It also finds applications in designing hearing aids where frequency lowering is customary. Recent developments in speech processing make use of artificial neural networks rather than the traditional rule-based approach. However, it is challenging as it demands huge labeled data. Unsupervised approach can overcome the laborious labeling task. Auto resonance networks are one choice for unsupervised approach. Recently, auto resonance networks have been used in classification problems. This work proposes an unsupervised approach for fricative detection using auto resonance networks. A system is built to detect the fricatives in both controlled speech and pathological speech. The system is tested using the standard TIMIT database for controlled speech and TORGO database for pathological speech utterances. The results are compared with another standard unsupervised neural network approach namely, Self Organizing Maps.