In recent years, the speech recognition has made significant progress because deep networks are utilized. But because of the noise and reverberation, the performance of far-field speech recognition is still unsatisfactory. Though weighted prediction error (WPE) with deep neural network (DNN) can reduce the noise and attenuate reverberation, it still has some shortcomings. In our work, we use the recurrent neural network with long short term memory (LSTM) to predict the coefficients in WPE, which makes up the disadvantages and gets a better performance in the speech recognition. The experiment results on CHiME-5 dataset show that the best model with the proposed method gains 2.1% absolute word error rate (WER) reduction compared to the baseline system.
Paper
Full text
Long Short Term Memory Networks Weighted Prediction Error for Far- Field Speech Recognition
Semantic Scholar · Computer Science · 2019
Abstract
In recent years, the speech recognition has made significant progress because deep networks are utilized. But because of the noise and reverberation, the performance of far-field speech recognition is still unsatisfactory. Though weighted prediction error (WPE) with deep neural network (DNN) can reduce the noise and attenuate reverberation, it still has some shortcomings. In our work, we use the recurrent neural network with long short term memory (LSTM) to predict the coefficients in WPE, which makes up the disadvantages and gets a better performance in the speech recognition. The experiment results on CHiME-5 dataset show that the best model with the proposed method gains 2.1% absolute word error rate (WER) reduction compared to the baseline system.