Spectral normalization employing hidden Markov modeling of line spectrum pair frequencies

This paper proposes a spectral normalization approach in which the acoustical qualities of an input speech waveform are mapped onto that of a desired neutral voice. Such a method can be effective in reducing the impact of speaker variability such as accent, stress, and emotion for speech recognition. In the proposed method, the transformation is performed by modeling the temporal characteristics of the line spectrum pair (LSP) frequencies of the neutral voice using hidden Markov models. The overall approach is integrated into a pitch synchronous overlap and add (PSOLA) analysis/synthesis framework. The algorithm is objectively evaluated using a distance measure based on the log-likelihood of observing the input (or normalized input) speech given Gaussian mixture speaker models for both the input and desired neutral voice. Results using the Gaussian mixture model formulated criteria demonstrate consistent normalization using a 10 speaker database