Biologically Based Non-Language Speech Sound Detection

This proposal addresses the application of new acoustic processing technologies to automatically identify and eliminate non-language speech sounds as a pre-processing stage to improve audio processing. Non-language speech sounds (coughing, breathing, ‘ah,’ ‘uhmm’) make up a large part of natural human language use, but contemporary speech recognition data preparation relies on hand-labeling of non-language speech sounds. The proposed work will extend, improve, and refine the capabilities of a computational model of auditory cortical processing based on multiscale spectro-temporal modulation features to the automated detection of non-language speech sounds. Advanced Acoustic Concepts and the University of Maryland have extensive experience applying the computational model to a variety of speech processing problems. Phase I results indicate that cortical processing algorithms are highly capable of identifying non-language speech sounds. Recognition algorithms will be trained and tested to distinguish non-language speech sounds collectively and individually, to classify them by individual type, and to determine accurate segmentation of the sound stream. Speech enhancement algorithms based on the same cortical processing model will be designed or adapted and applied to improve the identification of non-language speech sounds under noisy conditions. The feasibility of performing identification, classification, and segmentation in near real time will be demonstrated.