Audio Displays of Speech Segments at Three Levels of Phonetic Context

Abstract

Three computer programs to segment speech have been written for the Lincoln Laboratory TX-2 computer. The segments of these programs are at three levels of phonetic context. One is at a level below that of an individual speech sound, and corresponds to individual pitch periods in voiced sounds. A second is at the level of a speech sound or short sequence of speech sounds, and is correlated to a considerable extent with place and manner of articulation [J. Acoust. Soc. Am. 32, 1516 (1960)]. The third is above the level of single speech sounds, and is closely correlated with syllables. For the last two types of segments, the following technique is used so that the individual segments may be heard separately. The speech waveform is time- and amplitude-quantized and read into the computer. From the computer, it is stored on digital tape. The digitized waveform is again read out of the computer, decoded through a digital-to-analog converter, and fed into a 35-channel filter bank. The output of the filter bank is time- and amplitude-quantized and read into the computer. This spectral information is used to arrive at segment decisions. The quantized waveform is read back in from digital tape. The segment decisions are matched up against the waveform. The quantized waveform, from beginning to end of segment, is read out of the computer, decoded, and available for recording. For the segments of pitch period length, an individual pitch period may be read out over and over, giving a continuous audio display of its harmonic structure. Some of these outputs are demonstrated, and possible uses of this display facility are mentioned.