The phase vocoder is an analysis-synthesis system which has as intermediate data the time-variant discrete Fourier spectrum of the input signal. It can be formulated in a way such that the synthesized signal is identical to the original, both theoretically and practically. The intermediate data can be transformed, also with no loss of information, into the more conventional magnitude and frequency representation. This intermediate data can then be used to resynthesize the tone at different pitches or different rates than the original with the advantage that when no modification is made, the synthetic tone is absolutely identical to the original. This represents a significant advance over the Heterodyne filter, which placed severe restrictions on the amount of variation in pitch or amplitude that could be analyzed. The phase vocoder has no such restrictions and can just as easily deal with vibrato and inharmonic tones. Since calling the frequencies in the phase vocoder analysis data also scales the spectrum up, use of this modification with speech can produce altered vowel tones. If this method is combined with the linear predictor, using the phase vocoder to alter the pitch of the error signal, then the spectrum can be held constant while the pitch is changed, thus allowing independent control over time, pitch, and spectrum. Vowel quality can be preserved or altered at will. Again, if no modification is made, the combination of the linear predictor and the phase vocoder is an identity, both theoretically and practically. This search is still in a very preliminary state, so only a few sound examples can be given at this time. A full theoretical explanation, however, can be given.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.