Search

The paper reports the results of a digital computer simulation in which a sample of connected speech was analyzed and resynthesized in terms of a series of orthogonalized exponentially damped sinusoids. It was found to be possible to synthesize each pitch period from a function set having only 16 fixed frequencies with fixed damping at each frequency. The processing was done pitch‐synchronously on an IBM‐7090 digital computer, using 10‐bit accuracy speech which had been digitized at a rate of 12000 samples per second. Procedures for generating the sample values of the function set on a digital computer, manual pitch extraction, and data‐processing methods are described. Comparisons of the input and resynthesized speech samples are presented both visually and aurally by means of a tape recording. The results indicate that both the phonetic content and the quality of the speaker's voice are retained in the analysis‐synthesis process. This analysis‐synthesis process, simulated in this investigation, is amenable to real‐time analog realization. (This work was supported by the Rome Air Development Center under contract No. AF 30(602)‐2446.)

Time‐domain speech processors, that analyze the short‐time autocorrelation function in terms of a set of orthogonal functions, are shown to yield parameters that are the coefficients of a generalized Fourier cosine series representation of the speech power spectrum. Two examples are described, one using a Laguerre filter and one using a tapped delay line as in the autocorrelation Vocoder. In the case of the autocorrelation Vocoder, the derived power spectrum representation reduces to a simple Fourier cosine series in the frequency domain.

This paper describes a digital cepstrum‐pitch detector designed to work in real time on a 12‐bit binary machine. Spectrum and cepstrum analysis of raw speech data is accomplished through the use of a single complex Fast Fourier Transform (FFT) algorithm. Pitch detection is accomplished by a cepstrum‐peak scanning logic designed to locate the position of a peak corresponding to the pitch period. Since other peaks may exist in the data, logic is required to differentiate these peaks from the desired one. Because of the limited word size of the machine, special processing algorithms were developed to perform logging and to avoid the squaring of variables. Various logical operations performed on the spectrum and cepstrum functions in voicing detection and in the extraction of pitch are presented and discussed. A demonstration of the performance of this pitch detector when operated with a 2400‐bit/sec digital vocoder will be presented by means of a tape recording. [Work supported under contract.]

In this paper, an experiment is described in which articulation scores were measured as the number of transmitted spectral parameters was varied from 6 to 16 in a Laguerre autocorrelation vocoder. The transmitted spectral parameters were the expansion coefficients of the short‐time‐averaged speech autocorrelation function in terms of Laguerre functions. Thus, these parameters are the same as the expansion coefficients of the speech power spectrum in terms of the Fourier cosine transforms of the Laguerre functions. The analysis and synthesis filters in this vocoder were synthesized from a number of identical RC transistor circuits. The time constant of the Laguerre analysis and synthesis filters was 125×10−6 sec. The system was operated in an analog back‐to‐back mode for these tests. The results presented show that about 16 parameters are required to achieve performance comparable to contemporary filter‐bank vocoders. The intelligibility tests were performed by Bolt Beranek and Newman Inc., in early 1965.

Speech compression systems are described which make use of orthonormal function expansions of either the speech waveform itself or the short‐time autocorrelation function of the speech waveform. A system of this type theoretically can be optimized so that a minimum number of independent, slowly varying parameters are transmitted for a given level of desired performance. In both types of systems, the compression in data rate is obtained through the optimization of the expansion series, the truncation of the series, and by averaging the coefficients of the expansion over several pitch periods. In addition, the phase spectrum of the speech signal can be eliminated in a systematic manner in both the system wherein the waveform is analyzed and the systems where the autocorrelation function is analyzed.