sound source

source separation

speech enhancement

speaker recognition

speech signal

modulation recognition

mutual coupling

IEEE Journal of Selected Topics in Signal Processing

Supervised learning based methods for source localization, being data driven,\ncan be adapted to different acoustic conditions via training and have been\nshown to be robust to adverse acoustic environments. In this paper, a\nconvolutional neural network (CNN) based supervised learning method for\nestimating the direction-of-arrival (DOA) of multiple speakers is proposed.\nMulti-speaker DOA estimation is formulated as a multi-class multi-label\nclassification problem, where the assignment of each DOA label to the input\nfeature is treated as a separate binary classification problem. The phase\ncomponent of the short-time Fourier transform (STFT) coefficients of the\nreceived microphone signals are directly fed into the CNN, and the features for\nDOA estimation are learnt during training. Utilizing the assumption of disjoint\nspeaker activity in the STFT domain, a novel method is proposed to train the\nCNN with synthesized noise signals. Through experimental evaluation with both\nsimulated and measured acoustic impulse responses, the ability of the proposed\nDOA estimation approach to adapt to unseen acoustic conditions and its\nrobustness to unseen noise type is demonstrated. Through additional empirical\ninvestigation, it is also shown that with an array of M microphones our\nproposed framework yields the best localization performance with M-1\nconvolution layers. The ability of the proposed method to accurately localize\nspeakers in a dynamic acoustic scenario with varying number of sources is also\nshown.\n

Multi-Speaker DOA Estimation Using Deep Convolutional Networks Trained With Noise Signals