An Algorithm for Accurate Breath Detection in Speech and Song Signals

Automatic and reliable detection of pre-defined events in speech and audio signals is of great importance in many applications, and has been the subject of extensive research in recent years. One such application is in professional voice recordings, where the recorded signal may contain unwanted sounds or effects, and an automatic tool that could detect and manipulate these sounds is highly desirable. In this study we present an effective algorithm for detection of breath sounds in speech or song signals, in order to improve the aesthetics of the recorded voice. The algorithm works by creating a template feature matrix from the mel-frequency cepstral characteristics of several breath examples, and comparing it to feature matrices of consecutive frames of the audio signal, using an adaptive distance threshold, marking each frame as breathy or non-breathy. The initial detection is later refined by an edge detection algorithm, based on various waveform parameters, designed to demarcate the exact boundaries of each breath event and to eliminate possible false detections. Evaluation of the algorithm on a database of speech and songs containing several hundred breath sounds yielded a correct identification rate of 94%, with a specificity of 96%.