Sound Event Separation and Classification in Domestic Environment using Mean Teacher

Sound intelligence is gaining major interest of research and development community over the last couple of years. Sound event detection (SED) and classification result with other modalities such as vision, sensors etc. is helping in development of many new use cases. Overlapping sound events classification (SEC) in domestic environment is a challenging research task. Domestic environment includes sound signals from human, animals, devices and utilities. Overlapping of these signals reduces performance of SEC systems. This paper proposes sound source separation as preprocessing block for SEC system. Time-dilated convolutional network (TDCN) mask network with Shorttime Fourier Transform (STFT) is used for sound source separation. Log mel features are extracted, then mean teacher model with context gating Convolutional Recurrent Neural Network (CRNN), is trained for SEC system. The paper concludes that, the source separation as preprocessing and fusion of classification results on each separated signal using SEC, helps in achieving enhanced accuracy.