Speech Enhancement for Cochlear Implant Recipients using Deep Complex Convolution Transformer with Frequency...

Area of Science:

• Signal Processing
• Machine Learning
• Auditory Neuroscience

Background:

• Cochlear implant (CI) users face communication challenges in noisy environments due to distorted speech signals.
• Existing speech enhancement (SE) methods often neglect the crucial role of phase information in speech perception.

Purpose of the Study:

• To develop a novel deep learning model for simultaneous enhancement of speech magnitude and phase spectra.
• To improve speech intelligibility and quality for CI users in complex acoustic environments.

Main Methods:

• A Deep Complex Convolution Transformer Network (DCCTN) was proposed, utilizing a complex-valued U-Net with a transformer in the bottleneck.
• The network incorporates a frequency transformation block to capture speech harmonic correlations.
• DCCTN learns a complex transformation matrix for time-frequency domain speech recovery.

Main Results:

• DCCTN outperformed existing SE models (CRN, DCCRN, GCRN) in objective speech intelligibility and quality metrics.
• Formal listener evaluations with CI recipients confirmed significant improvements in noisy conditions.
• The model effectively suppressed non-stationary noise without introducing musical artifacts.

Conclusions:

• The proposed DCCTN offers a superior approach to speech enhancement for CI users by addressing both magnitude and phase distortions.
• This method holds promise for enhancing real-world communication for individuals with hearing impairments.