An extended Levinson-Durbin algorithm and its application in mixed excitation linear prediction

Area of Science:

• Speech Processing
• Digital Signal Processing
• Acoustics

Background:

• Linear predictive coding (LPC) is crucial for modeling the human vocal tract.
• The Levinson-Durbin algorithm efficiently solves Yule-Walker equations for LPC model coefficients.
• Optimizing LPC algorithms can enhance speech synthesis quality, particularly at lower bitrates.

Purpose of the Study:

• To develop an extended Levinson-Durbin algorithm for improved vocal tract modeling.
• To enhance the performance of 2400 bit/second mixed excitation linear prediction (MELP).
• To investigate the impact of modified iteration steps on speech quality.

Main Methods:

• A ten-order all-pole model was employed to represent the human vocal tract.
• An extended Levinson-Durbin algorithm was proposed with adjustable iteration steps.
• A hybrid approach combining the extended and traditional algorithms was applied to MELP.

Main Results:

• The extended algorithm accounts for interactions between adjacent subtractors in the Teoplitz matrix.
• The hybrid algorithm successfully solved the Yule-Walker equations under specific conditions.
• A noticeable improvement in the Mean Opinion Score (MOS) for nasal syllables was observed.

Conclusions:

• The proposed extended Levinson-Durbin algorithm offers a viable enhancement for LPC.
• The modified algorithm shows potential for improving speech quality in low-bitrate applications.
• Further research can explore broader applications of this enhanced algorithm in speech technology.