Reprogramming Automatic Speech Recognition Models for Neonatal Chest Sound Separation

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Stethoscope-recorded chest sounds are crucial for non-invasive cardiac and pulmonary assessments.
• Noisy chest sounds compromise the accuracy of diagnostic algorithms, necessitating robust preprocessing techniques.
• Current methods for chest sound separation often require complex preprocessing steps.

Purpose of the Study:

• To investigate the efficacy of reprogramming automatic speech recognition (ASR) models for neonatal chest sound separation.
• To evaluate two distinct reprogramming strategies for the Whisper ASR model: audio encoder-only and full model reprogramming.
• To demonstrate the potential of parameter-efficient model reprogramming for biomedical signal processing.

Main Methods:

• Reprogrammed the Whisper ASR model, a large foundational model, for chest sound separation tasks.
• Implemented two approaches: modifying only the audio encoder and reprogramming the entire model.
• Utilized simple linear layers and learnable parameters for efficient model adaptation.
• Tested the methods on an artificial dataset for neonatal chest sound separation.

Main Results:

• The parameter-efficient reprogramming of Whisper effectively separated heart and lung sounds from noise.
• The proposed method, when used as a preprocessing step, achieved performance comparable to state-of-the-art algorithms.
• Demonstrated successful application of a pre-trained ASR model for cross-domain biomedical sound separation.

Conclusions:

• Reprogramming ASR models offers a feasible and efficient method for neonatal chest sound separation.
• This approach highlights the potential of leveraging large, pre-trained foundational models in diverse scientific domains, including biomedical data.
• The study validates the effectiveness of cross-domain model reprogramming for enhancing diagnostic capabilities in healthcare.