Source separation for target enhancement of food intake acoustics from noisy recordings

Automatic food intake monitoring can be significantly beneficial in the fight against obesity and weight management in our society today. Different sensing modalities have been used in several research efforts to accomplish automatic food intake monitoring with acoustic sensors being the most common. In this study, we explore the ability to learn spectral patterns of food intake acoustics from a clean signal and use this learned patterns for extracting the signal of interest from a noisy recording. Using standard metrics for evaluation of blind source separation, namely signal to distortion ratio and signal to interference ratio, we observed up to 20dB improvement of separation quality in very low signal to noise ratio conditions. For more practical performance evaluation of food intake monitoring, we compared the detection accuracy for chew events on the mixed/noisy signal versus on the estimated/separated target signal. We observed up to 60% improvement in chew event detection accuracy for low signal to noise ratio conditions when using the estimated target signal compared to when using the mixed/noisy signal.