Multi-objective non-intrusive hearing-aid speech assessment model

Area of Science:

• Speech Processing
• Hearing Aid Technology
• Machine Learning

Background:

• Reference-free speech assessment is crucial due to the absence of reference signals in real-world applications.
• Existing deep learning models show promise but have limited focus on hearing-impaired (HI) subjects.
• Non-intrusive speech assessment is vital for numerous speech processing applications.

Purpose of the Study:

• To present HASA-Net+, an improved multi-objective, non-intrusive speech assessment model for hearing aids.
• To enhance speech quality and intelligibility prediction for both normal-hearing and HI listeners.
• To evaluate the model's robustness and generalization capabilities in diverse acoustic conditions.

Main Methods:

• HASA-Net+ builds upon the previous HASA-Net model, incorporating pre-trained speech foundation models and fine-tuning.
• The model's predictive capabilities were expanded to include various conditions: noisy, denoised, reverberant, dereverberated, and vocoded speech.
• Generalization was validated using an out-of-domain dataset.

Main Results:

• HASA-Net+ demonstrates improved performance in predicting speech quality and intelligibility.
• The model proves robust across diverse acoustic environments, including noise and reverberation.
• Validation with an out-of-domain dataset confirms the model's generalization capability.

Conclusions:

• HASA-Net+ offers a robust and inclusive solution for non-intrusive speech quality and intelligibility assessment, particularly for hearing aid users.
• The integration of foundation models and expanded testing conditions enhance its applicability.
• This model advances speech processing for hearing assistance and diverse acoustic scenarios.