Analysis of Speech Signals Using Excitation Source Information

Abstract

Speech is output of the time varying vocal tractsystem excited with the time varying excitation. Speech isproduced due to the impulse like excitation in each glottal cycle. During the production of speech, the instant of significantexcitation of the vocal tract system is referred to as epoch. In caseof voiced speech, most significant excitation takes place at theinstants of glottal closure i.e. glottal closure instants can bereferred as instants of significant excitation. Speech laugh is asignal produced when laughter occurs with neutral speech. Thespeech-laugh signal occurs frequently in natural conversationwith people. The features of speech-laugh, laughter and singingvoice deviates from the features of neutral speech. In this paper, we discriminate laughter, speech-laugh and neutral speech anddiscriminate singing voice and speech by obtaining epochlocations and extracting new features from these epochs. Themethod used here for the extraction of epochs is the ModifiedZero Frequency Filtering method. The features extracted fromepochs for the discrimination are fundamental frequency(f0) andslope of f0(α) at epoch locations and number of epochs (k) andstrength of excitation (β).