Single-Ended Speech Quality Prediction Based on Automatic Speech Recognition

摘要

A new single-ended speech quality measure is proposed that uses a deep neural network (DNN)-based automatic speech recognition system. A quality measure is used to quantify the degradation of the DNN output (phoneme posterior probabilities or posteriorgrams) caused by speech distortions. The new method was evaluated using five databases containing nine subsets of data covering several conditions of narrowband and broadband speech, distorted by speech codecs, telecommunication networks, clipping, chopped speech, echoes, competing speakers, and additional background noises. Since our model is trained as a speaker-independent speech-specific system, it is not suited for predicting speech quality in the presence of a background speaker. The evaluation results of all remaining eight data subsets show that good average correlations with subjective speech quality ratings are achieved without any task-specific training or optimizations (r = 0.81). These average results are close to those achieved with the American National Standard ANIQUE+ (r = 0.83) and clearly better than those obtained with the ITU-T standard P.563 (r = 0.58).