Estimation and Voicing Detection With Cascade Architecture in Noisy Speech.

Area of Science:

• Speech Processing
• Acoustics
• Machine Learning

Background:

• Pitch tracking is crucial for speech processing but challenging in noisy conditions.
• Non-stationary noises degrade speech signals, complicating pitch estimation and voicing detection.
• Accurate voicing detection is essential for reliable pitch tracking.

Purpose of the Study:

• To propose a neural cascade architecture for joint pitch estimation and voicing detection.
• To enhance pitch tracking performance in noisy speech.
• To develop a speaker-independent and noise-independent method.

Main Methods:

• A neural cascade architecture was designed, integrating speech enhancement and pitch tracking modules.
• The multi-task framework was trained to optimize both pitch estimation and voicing detection simultaneously.
• The model was trained in a speaker-independent and noise-independent manner.

Main Results:

• The integrated enhancement module significantly improved pitch estimation and voicing detection accuracy, particularly at low signal-to-noise ratios (SNRs).
• The proposed multi-task framework outperformed combined single-task models in performance and efficiency.
• The method demonstrated robustness across various noise conditions.

Conclusions:

• The neural cascade architecture offers a superior and efficient approach to pitch tracking in noisy speech.
• Joint optimization of speech enhancement and pitch tracking is effective for improving accuracy.
• The proposed method shows significant advantages over existing pitch tracking techniques.