Optimizing pulse amplitudes in multipulse excitation

Abstract

Multipulse excitation [B. S. Atal and J. R. Remde, Proc. ICASSP 82] is an attractive way of modeling the input to the LPC synthesis filter at low bit rates. In the method described by Atal and Remde, the amplitudes and locations of the pulses are determined in successive stages—one pulse at a time. At each stage, one pulse is determined to achieve maximum reduction in the weighted mean-squared error between the original and synthetic speech waveforms. For closely spaced pulses, successive optimization of individual pulses is inaccurate, and often requires additional pulses to compensate for inaccuracies introduced earlier. This problem can be only partly corrected by reoptimizing all pulse amplitudes at the end. This paper describes a method which locates s pulse at any stage by jointly optimizing its amplitude together with the amplitudes of all pulses located up to that stage. As a result the amplitudes of all pulses are kept optimal at every stage, and very little inaccuracy is introduced even if the pulses are closely spaced. The new method provides significant improvements in the SNR—by as much as 10 dB in some sections of speech.