Transient-based speech transmission index for predicting intelligibility in nonlinear speech enhancement processors

A new speech intelligibility metric is proposed for the assessment of speech enhancement processors. These processors usually affect the fine structure in speech that is of fundamental importance to speech intelligibility. Classical metrics analyze the entire signal and thereby generally overestimate intelligibility. The measure presented here, therefore, isolates speech-transients by a cepstral smoothing technique and subsequently calculates speech intelligibility using an efficient version of the speech transmission index. By means of a genetic optimization of adjustable parameters, the proposed transition-based speech transmission index (TB STI) is adapted to the subjective data of linearly and nonlinearly processed speech. The method was assessed on untrained subjective data and showed a considerable improvement over other well-established measures.