A frame selective dynamic programming approach for noise robust pitch estimation

Area of Science:

• Speech processing
• Signal analysis
• Acoustics

Background:

• Existing pitch estimation techniques offer complementary approaches.
• Sub-harmonic to harmonic ratio (SHR) and sawtooth-wave inspired pitch estimator (SWIPE) are prominent methods.
• Combining complementary methods can potentially improve pitch estimation accuracy.

Purpose of the Study:

• To propose a novel frame selective dynamic programming (FSDP) method for pitch estimation.
• To exploit the complementary characteristics of SHR and SWIPE.
• To enhance the accuracy and robustness of pitch detection in speech signals.

Main Methods:

• A frame selective dynamic programming (FSDP) approach is proposed.
• FSDP classifies voiced frames into two classes for distinct processing.
• It utilizes variants of SHR and SWIPE, incorporating dynamic programming (DP) for selected frames.

Main Results:

• The FSDP method demonstrated superior pitch estimation performance compared to baseline schemes across various signal-to-noise ratio (SNR) conditions.
• FSDP achieved lower voice/unvoice (VuV) classification error than all four baseline schemes.
• The effectiveness was validated on KEELE, CSLU, and PaulBaghsaw corpora.

Conclusions:

• The proposed FSDP method offers significant improvements in pitch estimation accuracy.
• Applying a smoothness constraint using DP on selected frames enhances performance.
• FSDP provides a more robust and accurate solution for pitch detection, even in noisy conditions.