Bird Call Classification Using DNN-Based Acoustic Modelling

Bird call recognition using deep neural network-hidden Markov model (DNN-HMM)-based transcription is proposed. The work is an attempt to adapt the human speech recognition framework for bird call classification through transcription approach. Initially, the phone transcriptions are generated using CMU-Sphinx, and lexicons are modified using group delay-based segmentation. Later, bird call transcription is implemented using hybrid DNN-HMM framework through DNN-based acoustic modelling. During the DNN-based acoustic modelling, mel-frequency cepstral coefficient features (MFCCs) are computed and experimented with monophone models, triphone models, followed by linear discriminative analysis and maximum likelihood linear transform. The transcribed phonemes are corrected using context-based rules in the final phase. The proposed approach is evaluated on a dataset that consists of ten species with 563 audio tracks. The hybrid DNN-HMM approach outperforms the convolutional neural network and long short-term memory framework with an accuracy of 94.46%.