A Robust and Low Computational Cost Pitch Estimation Method.

Area of Science:

• Signal Processing
• Acoustics
• Speech Technology

Background:

• Accurate pitch estimation is crucial for speech and audio signal processing.
• Existing methods struggle with the non-ideal harmonic structure of real-world signals due to vocal tract characteristics, equipment acoustics, and spectral leakage.
• Harmonic frequencies often deviate slightly from integer multiples of the fundamental pitch.

Purpose of the Study:

• To enhance the accuracy and robustness of pitch estimation methods.
• To address limitations in modeling harmonic structure for pitch detection.
• To improve computational efficiency in pitch estimation algorithms.

Main Methods:

• Modified the Summation of Residual Harmonics (SRH) method.
• Relaxed the strict integer multiple constraint for spectral peak positions to allow slight deviations.
• Introduced a low-computational cost pitch segment extension scheme, integrated into voiced/unvoiced decision for error reduction.

Main Results:

• The proposed method demonstrated superior overall accuracy and robustness compared to ten other pitch estimation techniques.
• Experimental results showed a significant reduction in time cost, approximately 1/8th of the fast Non-Linear least Squares (NLS) method.
• Improved capture of harmonics and more efficient utilization of pitch smoothness prior.

Conclusions:

• The modified SRH method with pitch segment extension offers a more accurate and robust solution for pitch estimation.
• The approach effectively handles deviations in harmonic frequencies, improving performance on real-world speech and audio signals.
• The method provides a substantial speed advantage, making it suitable for computationally constrained applications.