Using cepstral coefficients for Inhalation pause detection in spontaneous speech

TL;DR: The method is most suited for signals with low noise and high average intensity (studio recording) but can also be used on noisier recordings with lower average intensity, albeit with poorer results.

Abstract: A method for recognizing inhalations in spontaneous speech is presented. It is similar to the template matching technique; a distance measure is calculated between a reference sound and an equally long portion of the same sound being tracked. A feature representation consisting of the standard Mel Frequency Cepstral Coefficients (MFCC), obtained by performing a discrete Cosine Transform of the mel-scaled filterbank spectrum is used. MFCC's are calculated every 5 ms. The comparison is then done by computing the euclidian distance between the cepstral coefficients of each frame of the two sounds. A low distance value means that the two compared inhalations are likely to be similar. The method can detect inhalations in both male and female spontaneous speech. The method is most suited for signals with low noise and high average intensity (studio recording) but can also be used on noisier recordings with lower average intensity, albeit with poorer results. (Less)