Hybrid BiLSTM-HMM based event detection and classification system for food intake recognition

We present a hybrid approach of chew sound event identification grounded on a Bidirectional Long Short-Term Memory plus Hidden Markov Model hybrid system in this paper (Bi-LSTM-HMM). We use a state-of-the-art chewing recognition recurrent neural network (RNN) and HMM model to the multi-label classification problem. In comparison to the basic implementation of convolutional neural network (CNN), Long Short-Term Memory (LSTM) and BiLSTM model this modification includes an unambiguous temporal model for output labels. Using data from an in-house lab developed data acquisition system iHearken, the efficacy of our proposed approach is compared to existing techniques. Our proposed model surpassed traditional models both in event detection and food classification, percentage of frames classified correctly of 92.6 %, and average Equal Error Rate (EER) of 12.51