Flexible and efficient harmonic resynthesis by modulated sinusoids

An harmonic plus noise system is presented that uses interpolation across smooth harmonic and noise spectral surfaces as an alternative to copying the spectral envelope from analysis instants. This combines the benefits of isolating the harmonic and noise parts of speech, and simplifying the application of complex transformations to the vocal tract filter. This has been coupled with a method of resynthesising the harmonic part of speech based on modulated sinusoids that more accurately models the harmonic component, and is half the time complexity of traditional overlap and add synthesis. Both the general case of synthesising across arbitrary voiced segments, and a simpler case for synthesising between pitch synchronous instants are detailed. The necessity of two dimensional phase unwrapping is discussed, and a solution presented. This approach is shown to result in high quality modified speech in a perceptual test comparison with the STRAIGHT system.