GB1487291A - Determining the boundaries of a speech utterance - Google Patents

Abstract

1487291 Speech boundary detectors WESTERN ELECTRIC CO Inc 24 March 1975 [29 March 1974] 12245/75 Heading H4R The beginning and/or the end of a speech utterance is determined by adaptively encoding the speech (Fig. 1) and determining when the energy of the encoded speech signals exceeds a threshold. The speech input is compared with the level predicted from the quantized level in loop 14, 15 to produce a differential signal input to the quantizer 12 whose output, at e.g. 16 levels, is encoded 13 to produce a series of digital output code words C(i). The quantizing steps are automatically adjusted to take account of the signal variations by logic network 16. It is said that the code word energy, which is equivalent to the number of code word adaptations per unit time, indicates far better than the speech waveform energy, where the boundaries of an utterance lie. Voiced and unvoiced utterances show no significant difference in code word energy. An expression of this energy E(n) is taken as the sum over a window of 101 sampling instants of the square of the individual code components the average codeword energy is subtracted from each C(i) before evaluation of the energy expression. The D.C. level here is 7À5 but since this cannot be expressed easily in digital form, the terms C(i) are replaced by a(i) = [2C(i) - 15]<SP>2</SP> in the expression for E(n) and this leads to the equation E(n) = E(n - 1) + a(n+50)-a(n-50), where a(n + 50) and a(n - 50) represent the newest and oldest values of a. The circuit of Fig. 4 performs this computation to generate E(n) which then passes to a logic circuit (Fig. 5, not shown) which compares E(n) with a stored threshold between the values of E associated with "silence" and average measured speech. When the threshold is first exceeded a counter begins to increment and continues for as long as the threshold remains exceeded until a count of 320 samples, equivalent to 50 Ásecs. of speech, triggers an output pulse confirming that the original input is speech. Similarly when the threshold is crossed in the downward direction, a count of 1024 (160 Ásec.) produces an output confirming that speech has ceased. These outputs may be used to indicate to an operator that a speech utterance has occurred or to gate a register temporarily storing the code words before permanent storage without gaps between utterances.