Time-frequency processing of partials for high-quality speech synthesis

摘要

Based on the particularities offered by the chosen signal model, we introduce a novel approach regarding the chain of actions pursued in the analysis stage of the speech signal, which succeeds the level of partial extraction. According to the harmonic plus noise model (HNM) a number of successive estimation and synthesis operations are performed. The present paper proposes a method to enhance the harmonic parameters estimation. This new algorithm proves to have good behavior offering support in selecting an appropriate subset of partials. In addition to reducing the arithmetic complexity of the harmonic synthesis (which is known to be the most resource consuming module), this optimized selection of the partials allows us to perform a specific frequency correction, which enables the possibility of a simple and coherent future pitch manipulation. The performed experiments confirmed the expected complexity reduction. Moreover, we applied the proposed algorithm for partial selection and tracking followed by a frequency aligning of the harmonic components. The reconstructed signal compared to the original speech proved to be a perceptually indistinguishable replica.