Evaluation of a Statistical Approach to Voiced-Unvoiced-Silence Analysis for Telephone-Quality Speech

In a recent paper, Atal and Rabiner described a fairly sophisticated method for reliably classifying segments of a waveform as voiced speech, unvoiced speech, or silence.1 The analysis method used a statistical pattern-recognition approach to make this three-class decision. In another investigation, Rabiner et al. showed that the accuracy of the classification algorithm was quite high when the input signal was wideband; however, for telephone speech inputs, the accuracy of the classification degraded quite significantly. 2 The reason for this result 455 was not that the method inherently broke down for telephone inputs, but instead that the particular parameter set effective for wideband inputs was not equally effective for band-limited inputs. Thus, the motivation for the work to be presented in this paper is to investigate the suitability of a large number of parameters as features for reliable voiced-unvoiced-silence classification for telephone-quality speech. Figure 1 shows a block diagram of the basic voiced-unvoiced-silence analysis algorithm. As shown in this figure, there are three steps in the method. First the speech is preprocessed. Generally, this preprocessing is a simple filtering operation; e.g., in the earlier work, a 200-Hz highpass filter was used to remove dc, hum, or low-frequency noise components present in the input signal. For telephone line inputs, we have considered somewhat more sophisticated preprocessing; namely, we have studied the use of a second-order inverse filter (as originally proposed by Itakura3) to normalize out the effects of varying telephone lines.