Models of speech and hearing and their applications: Overview and outlook

Abstract

Electronic speech coding had its origin in 1928 when Homer Dudley proposed to send speech signals over a new transatlantic telegraph cable with the (then) enormous bandwidth of 100 Hz. Dudley argued that speech was generated by slowly moving articulators and should therefore require only some 100-Hz total bandwidth for transmission. Since extraction of these parameters proved difficult, Dudley suggested, as an alternative, the transmission of the likewise slowly changing spectral information. The result was the channel vocoder, which (together with formant and voice-excited vocoders) dominated speech coding and that of musical sounds of single-pitch instruments until the mid 1960s. Unhappy with the vocoder's “electronic accent.” Bishnu Atal and the author, in 1967, advanced the idea of linear predictive coding (LPC), motivated by the desire to replace rigid vocoderlike coding by a more flexible method in which a “prediction residual” could represent deviations of real speech signals from an assumed all-pole model. In 1972, subjective error criteria were introduced, based on auditory models of masking and the possibility of “hiding” quantizing noise under the speech signal. The resulting code-excited linear predictive (CELP) coders based on statistical models have been operated at 1/4 bit per sample for the residual at virtually no audible degradation. With powerful digital signal processing chips, CELP has now become the method of choice for synthetic-voice applications. For musical sounds, masking models and linear predictive coding, which is equivalent to prewhitening of the spectrum, lead to a substantial reduction in the required number of bits per sample.