Privacy-Preserving Deep Speaker Separation for Smartphone-Based Passive Speech Assessment.

Area of Science:

• Speech processing
• Machine learning
• Biomedical engineering

Background:

• Smartphones offer passive speech impairment monitoring for conditions like Parkinson's disease and Alzheimer's.
• Speaker separation is vital for isolating target speech from background noise and cross-talk in recordings.
• Privacy concerns necessitate using speech features like Mel-Frequency Cepstral Coefficients (MFCCs) instead of raw audio.

Purpose of the Study:

• To introduce Deep MFCC bAsed SpeaKer Separation (Deep-MASKS) for improved speaker separation.
• To enable accurate speech analysis from smartphone recordings while preserving speaker privacy.

Main Methods:

• Deep-MASKS employs an autoencoder and Deep Neural Network (DNN) to reconstruct MFCCs.
• Speech representations (i-vector, x-vector, d-vector) are used for MFCC component reconstruction.
• The method processes continuous audio recordings, unlike utterance-based approaches.

Main Results:

• Deep-MASKS achieved up to a 44% reduction in Mean Squared Error (MSE) for MFCC reconstruction.
• The method reduced the additional bits needed for clean speech entropy by 36%.
• Performance surpassed existing baseline methods.

Conclusions:

• Deep-MASKS provides a privacy-preserving and accurate method for speaker separation using MFCCs.
• This technique is suitable for passive speech monitoring applications on smartphones.
• The DNN-based reconstruction offers superior performance over previous masking techniques.