SNR-Based Assessment of Quality of Speech Enhancement Using Single-Channel Methods

Abstract:

In the paper, the authors discuss methods for assessing the quality of noise-degraded speech using the signal-to-noise ratio (SNR). This objective assessment is of advantage in the evaluation of speech encoding and also in the rough assessment of the effectiveness of single-channel methods of speech enhancement. The application of listening tests is costly and therefore subjective assessment methods are only used for optimized enhancement methods. First, the classical definitions of SNR are given, where the knowledge of undisturbed speech signals is assumed. Besides the classical definitions, the definitions of SNR for averaging finite signal frames of different energy levels are considered, as is the case of speech. The case is also discussed when only a mixture of useful signal and noise is available. Here it is no longer possible to use the classical definitions; it is necessary to identify noise in speech pauses. For this purpose, a suitable voice activity decoder (VAD) is necessary. The classical methods for assessing the quality of speech were mainly proposed and used to assess vocoders, in which it can be assumed that non-distorted speech signal is available. In the case of single-channel enhancement methods no uncorrupted speech signal is available and therefore the definition of SNR must be slightly adapted or new approaches must be sought.


Year: 2007
In session: Signalverarbeitung
Pages: 149 to 154