Polyphonic Sound Event Detection and Classification using Convolutional Recurrent Neural Network with Mean Teacher

Abstract

Polyphonic sound event detection (SED) is research field which finds usefulness in cognitive IoT, security systems, voice assistants etc. This paper proposes a SED system, which makes use of semi supervised mean teacher approach, for DESED dataset, which has confined strongly classified data, and plethora of weakly labeled and unlabeled data. The sound data is processed by extracting the log mel spectrograms. A convolutional recurrent neural network (CRNN) model, which takes the log mel spectrograms as input, is used in first level of SED system. Then, the second level updates it further by using the mean teacher approach by learning from the weakly classified and unlabeled audio data. The proposed SED system is proved to be effective on the DESED dataset. Overall performance assessment is made with baseline as well as different top-ranked solutions. Event based F1-score is considered as performance evaluation metric. The experiments display that the proposed SED system with CRNN using mean teacher approach achieves F1-score of around 48.3%, which is significant improvement over baseline system.