On Instrumental Quality Assessment of Speech Enhancement Systems in Three Independent Dimensions

Instrumental quality assessment serves as an important tool for evaluating and developing noise reduction systems. Commonly three measures are of major interest: the quality of the speech component, the level of noise attenuation, and the amount of musical tones, as a specific type of noise distortion. The first two measures can already be obtained in an instrumental manner in many applications [1], see also ITUT Recommendations P.1100 [2] and P.1110 [3]. Nevertheless, noise distortions such as musical tones are usually still being subjectively evaluated. Recently, a high correlation of the perceived level of musical tones with the log-kurtosis ratio has been observed in [4]. However, this measure is computed from the noisy speech signal and the enhanced speech signal, making this ratio dependent on the level of noise attenuation. In addition, [4] assumes that the noise and speech signals are gamma-distributed in the power spectral domain. Motivated by the idea of an instrumental measure for residual noise distortion being independent of speech distortion and noise attenuation, a modified approach to the log-kurtosis ratio measure is proposed in this paper. Furthermore, two other quantities are presented for evaluating the quality of the speech component and noise attenuation.