Retracted: Signal Quality Assessment for Speech Recognition using Deep Convolutional Neural Networks

Due to the fact that it does not require the actual reference voltage, the non-intrusive speech quality evaluation system has recently gained favour. About the same time, neural networks were successfully implemented for assessing the quality of spoken language. In order to enhance the precision of non-invasive speech quality assessment, a neural network-based evaluation method is proposed in this study. The systems described here are based on convolutional neural networks (CNNs), bidirectional long short-term memory (BLSTM), and a CNN-LSTM architecture. When we examine the theoretical and empirical performance of four distinct pooling functions, the attention pooling mechanism stands head and shoulders above the rest. Experiments make use of a dataset containing several degraded speech sounds and subjective performance scores. Results show that the proposed CNN model with attention pooling function outperforms the standardised ITU-T P.563 and the autoencoder-support vector regression technique, with an R-value and RMSE of 0.967 and 0.269, respectively. Our model was put to the test on both custom-created and publicly-available datasets featuring ambient noise. Analysis is performed using correlation measures and statistical summaries. After extensive testing, we determined the following: As expected, the presented model generally gives samples with pure speech high ratings and examples with merely noise poor scores; (1) it had relatively high generalisation despite having been trained with a small number of samples; (2) it also identified speaking and background noise even for datasets with complicated deterioration; (3) it performed well overall; and (4) it performed as expected, generally, overall.