Study and Analysis of Signal Display and Bandwidth Compression Techniques

摘要

Parameter signals were generated representing (1) short-time running-average rate of occurrence of threshold crossings for an input speech signal (Z sub 1) or for the derivative of input speech signal (Z sub 2); and (2) short-time running-average amplitude envelope of the input speech (A). Pairs of these parameter signals were plotted as x- and y-deflections on the display of a storage oscilloscope (the 'vector' display), for speech elements of syllabic length, and the resulting trace patterns ('vectorgrams') were studied. Gross structure of the vectorgram patterns shows a significant tendency to be phoneme-correlated and talker-invariant, is little influenced by accents or by male-female voice distinctions, but is quite sensitive to clarify of enunciation. Fine structure is also seen, some of which is also seen, some of which is also phoneme-correlated and voice-invariant, and which is often repeated with great exactitude when a given speaker repeats a syllable. The relation of these patterns to other work on formant changes is discussed, as is their application to automatic speech recognition and bandwidth compression. Experiments are described in which the A and Z sub 1 or Z sub 2 parameter signals were passed through narrowband channels and used to reconstruct synthetic speech in a simulated receiver. Questions of instrumentation are discussed, and a technique of system evaluation by direct comparison of input and output vectorgrams is described. Possibilities for improving performance are considered. (Author)