| Abstract: | In real applications, speech signal is
usually corrupted by background noise, which greatly affects the performance of speech
processing systems. This paper addressed the important problem of noisy speech endpoint
detection in robust speech recognition or other speech processing areas. As the traditional
energy based method for speech endpoint detection had poor poerformance in low signal to
noise ratio, a novel speech endpoint detector was proposed. This approach takes advantages
of a Recurrent Neural Network (RNN) to determine speech activity. The RNN, which is trained to
be insensitive to noise variation, shows good behavior in the test experiments. |
