Data-driven voice source waveform analysis and synthesis

A data-driven approach is introduced for studying, analyzing and processing the voice source signal. Existing approaches parameterize the voice source signal by using models that are motivated, for example, by a physical model or function-fitting. Such parameterization is often difficult to achieve and it produces a poor approximation to a large variety of real voice source waveforms of the human voice. This paper presents a novel data-driven approach to analyze different types of voice source waveforms using principal component analysis and Gaussian mixture modeling. This approach models certain voice source features that many other approaches fail to model. Prototype voice source waveforms are obtained from each mixture component and analyzed with respect to speaker, phone and pitch. An analysis/synthesis scheme was set up to demonstrate the effectiveness of the method. Compression of the proposed voice source by discarding 75% of the features yields a segmental signal-to-reconstruction error ratio of 13 dB and a Bark spectral distortion of 0.14.

► Voice source signal is segmented and analyzed. ► Data driven model of voice source using GMM and PCA. ► Voice source prototypes derived. ► 25% compression of the voice source achieved with analysis/synthesis