STI
Improvements of STI: frequency weighting, gender, level dependent masking, and phoneme specific prediction
Herman J.M. Steeneken and Tammo Houtgast
1. Introduction
In the early 1970’s STI was used in many applied research projects. This provided experience for application in many practical conditions and also showed us some restrictions. For example the validity of STI for systems with extremely limited bandwidth gave us the impression of underestimation of the intelligibility. We also recognised that extension with female speech for the prediction of the intelligibility was required. Therefore, in 1989, a study was started to cover these issues and in the course of this study some other issues were raised. The following topics were investigated:
(1) The original algorithms of STI (Speech Transmission Index, Steeneken and Houtgast, 1980) and AI (Articulation Index, Fletcher, 1953; Kryter, 1962) assume that the prediction of the intelligibility is based on a weighted contribution for a number of frequency bands. For each band the effective signal-to-noise ratio (SNR) is determined. The effect of ambient noise, temporal distortion (reverberation, echoes), non-linear distortion (distortion components), and auditory masking, determine this effective signal-to-noise ratio. STI and AI assume that the frequency weighting is independent of the SNR. However, this was never tested experimentally and therefore requires verification.
(2) Application of the STI during the 1980’s taught us that for some specific conditions, errors of the prediction of intelligibility by the STI occurred. This was especially the case for conditions that included gaps in the frequency transfer or in case of a very limited frequency transfer. The latter occurs with the use of small horn loudspeakers, which have a typical frequency response that begins around 1000 Hz (mainly due to a limited length of the horn). It was found that for some frequency regions redundancy between adjacent frequency bands of the speech signal occurred. This led to a revised model of the frequency-weighting algorithm for calculation of STI.
improvements over three well-known existing methods...The gain, Γm , is a weighting function that ...function of frequency, masker level, and signal ...
improvements over three well-known existing methods...The gain, Γm , is a weighting function that ...function of frequency, masker level, and signal ...
热门文档