Chapter 5 - Harmonic adaptive speech synthesis

Abstract

This chapter presents a novel approach to speech synthesis by adaptively constructing and combining the harmonic components. Harmonic adaptive speech synthesis foundations are based on the fusion of Fourier series and adaptive filtering. Speech is a well-known biometric signal composed of a linear combination of sine and cosine components, which under stationary condition, can be effectively presented as a Fourier series. Theoretically the Fourier series coefficients for each stationary speech segment can be obtained by corresponding integration equations over the segment’s length. This chapter presents a practical online adjustment of the coefficients by an adaptive harmonic filter (AHF) that tracks the real speech and adaptively forms the Fourier series coefficients thereby synthesizing the corresponding speech. The experiments are carried out on speech synthesis from real speech records using an AHF with a least means square learning algorithm. For a number of real speech segments, the synthesized speeches are generated by AHF. The results are numerically evaluated by the correlation values between the real segments and synthetic ones. The correlation-based confusion matrix indicates a high level of similarity of each synthetic speech segment to its corresponding real segment.