Multipulse-excitation applied to vocoders
Multipulse-excitation has greatly improved the speech quality achievable from linear predictive coders which previously required speech to be classified as voiced or unvoiced for excitation purposes. Multipulse removes the need for voicing classification, improving speech quality by enhancing the excitation and offsetting errors in the vocal tract filter. An investigation of multipulse-excitation applied to a channel vocoder and a formant synthesiser was conducted. The prime objective was to improve the performance of these algorithms and achieve multipulse linear prediction speech quality, our target quality. This dissertation outlines and restates the idea of multipulse-excitation applied to a linear predictive vocoder. We then examine a high quality channel vocoder and formant synthesiser, and the use of multipulse-excitation to improve their performances. In each case time and frequency domain multipulsecalgorithms were used. Various modifications were made to these algorithms in order to accommodate multipulse-excitation and improve the overall speech quality. In the case of the channel vocoder this involved a novel technique, which sacrificed the inherent waveform preserving properties of the multipulse algorithm. Only by increasing both the pulse rate and the number of channels could the multipulse-excited channel vocoder achieve our target quality. With the formant synthesiser it was possible, by variation of the pulse rate alone, to achieve our target quality. Comparisons are drawn between the three multipulse algorithms and reasons given for their differing performance; this is substantiated by experimental results. These results suggested interesting improvements to the multipulse-excited formant synthesiser; and also hinted at a new and novel technique for formant tracking, using multipulse-excitation applied to a formant synthesiser.