Optimization of RNN-Based Speech Activity Detection

Speech activity detection (SAD) is an essential component of automatic speech recognition systems impacting the overall system performance. This paper investigates an optimization process for recurrent neural network (RNN) based SAD. This process optimizes all system parameters including those used for feature extraction, the NN weights, and the back-end parameters. Three cost functions are considered for SAD optimization: the frame error rate, the NIST detection cost function, and the word error rate of a downstream speech recognizer. Different types of RNN models and optimization methods are investigated. Three types of RNNs are compared: a basic RNN, long short-term memory (LSTM) network with peepholes, and a coordinated-gate LSTM (CG-LSTM) network introduced by Gelly and Gauvain. Well suited for nondifferentiable optimization problems, quantum-behaved particle swarm optimization is used to optimize feature extraction and posterior smoothing, as well as for the initial training of the neural networks. Experimental SAD results are reported on the NIST 2015 SAD evaluation data as well as REPERE and AMI meeting corpora. Speech recognition results are reported on the OpenKWS'13 test data. For all tasks and conditions, the proposed optimization method significantly improves the SAD performance and among all the tested SAD methods the CG-LSTM model gives the best results.