Model based voice decomposition method with time constraint

This paper proposes a new voice decomposition method with time constraint. Speech recognition of mixture of two and more voices and sounds is still very difficult. The model-based voice decomposition method proposed in our previous study solves the above problem; however, the solution is of a local optimal problem and the given spectral sequence sometimes varies rapidly and is non-realistic behavior. A new decomposition method solves a global optimal problem and the given spectral sequence changes are milder due to the time continuity constraint. This paper formulates the decomposition problem as an optimal path searching in the time-frequency domain. As the result of evaluation experiments, the average decomposition distortion is 4.16 dB and about 0.92 dB improvement is achieved