Method and apparatus for speech analysis

Abstract

Speech analysis is performed by synchronizing, with the initiation of glottal pulses produced during the voicing of phonemes, a pitch-independent transform that performs analysis entirely within a pitch period. Analysis is made of the three most dominant formant frequencies of the sound sets involved in the enunciation of voiced phonemes. Peaks in the transform represent formant frequencies. Three peaks are selected from the transform identifying the frequencies with the first, second and third greatest amplitudes. Correlation of the waveform between successive pitch periods detects whether a sound is a vowel, a voiced sibilant, or unvoiced. Unvoiced sound sets are similarly analyzed but the analysis is synchronized with artificially generated synch pulses. The sound sets are analyzed to determine (1) whether the sound is a vowel, a voiced sibilant or an unvoiced sound; (2) the primary formant frequency, i.e. the one with the greatest amplitude; (3) the substantial presence of secondary and tertiary formant frequencies with the next two largest amplitudes, as well as their frequencies; (4) the frequency ratios of the primary to the secondary and the primary to the tertiary frequencies; (5) the decay rate of at least the greatest formant frequency; and (6) the time relationships of these sound sets to one another and to unvoiced sounds. A unique set of phoneme identifications is derived from these factors. Apparatus using these factors for phoneme encoding and speech reproduction from encoded phonemes is disclosed.