Pitch contour stylization using an optimal piecewise polynomial approximation.

Area of Science:

• Signal Processing
• Speech Technology
• Computational Linguistics

Background:

• Speech synthesis requires accurate pitch contour modeling.
• Existing approximation methods may introduce audible errors.
• Minimizing approximation error is crucial for natural-sounding speech.

Purpose of the Study:

• To introduce a novel dynamic programming (DP) approach for piecewise polynomial approximation of discrete data.
• To apply this DP technique for effective speech pitch contour stylization.
• To objectively and subjectively evaluate the performance of the proposed method.

Main Methods:

• Developed a dynamic programming algorithm for piecewise polynomial approximation.
• Minimized the L2 norm of the approximation error.
• Applied the method to stylize speech pitch contours.
• Conducted objective evaluations using Mean Square Error (MSE).
• Performed subjective listening tests for speech quality assessment.

Main Results:

• The DP-based method achieved the minimum Mean Square Error (MSE) compared to other approximation techniques.
• Objective evaluations confirmed the superiority of the DP approach in minimizing approximation error.
• Subjective evaluations indicated that synthesized speech quality using the DP-stylized pitch contour was nearly identical to the original speech.

Conclusions:

• The proposed dynamic programming method offers an effective way to stylize speech pitch contours.
• This technique significantly reduces approximation error, leading to high-quality synthesized speech.
• The DP-based approach provides a robust solution for speech processing applications requiring accurate pitch contour modeling.