Whisperization and Masked CycleGAN-Based Framework for Electrolaryngeal Speech Enhancement

Area of Science:

• Speech processing
• Biomedical engineering
• Artificial intelligence

Background:

• Electrolarynx (EL) speech rehabilitation for phonation disorders offers limited acoustic cues, impacting speech quality.
• Existing EL speech suffers from degraded quality due to reliance on external mechanical sources, hindering further enhancement.
• The need for improved acoustic and perceptual properties in EL speech is critical for better patient outcomes.

Purpose of the Study:

• To propose a novel speech enhancement framework for electrolarynx (EL) speech.
• To improve the naturalness, intelligibility, and acoustic properties of EL speech.
• To explore the combination of whisperization and Masked CycleGAN for dual-stage speech enhancement.

Main Methods:

• A novel framework combining whisperization and a Masked CycleGAN model was developed.
• Whisperization processed EL speech to remove mechanical noise and constant excitation, creating whisper-like EL (W-EL) speech.
• Masked CycleGAN with frame-level masking reconstructed missing prosodic and linguistic features in W-EL speech.

Main Results:

• Acoustic analysis showed W-EL speech spectrograms, fundamental frequency (F0) values, and F0 contours more closely resemble normal speech.
• The enhancement process compensated for missing low-frequency energy below 500 Hz.
• Objective and subjective evaluations confirmed significant improvements in speech naturalness and intelligibility compared to original EL speech.

Conclusions:

• The proposed whisperization and Masked CycleGAN framework effectively enhances EL speech quality.
• This dual-stage enhancement achieves both redundancy removal and deficiency compensation.
• The findings offer a new pathway for EL speech enhancement and insights for pathological speech improvement.