Voice activity detection in the presence of breathing noise using neural network and hidden Markov model

This paper proposes a voice activity detection algorithm to be used in the presence of breathing noise. We use a hybrid approach where a neural network is first applied in individual frames using mel-band energies within the frame as inputs. The output of the neural network is then processed using a hidden Markov model, which takes into account the temporally continuous nature of speech activity. Both the neural network and the hidden Markov model can be trained in supervised manner. On simulations with realistic acoustic material, the proposed method achieved average frame-level sensitivity above 97% and average specificity above 95%. The proposed algorithm enables a good rejection of noise and breathing frames while retaining the intelligibility of input speech.