11th Speech in Noise Workshop, 10-11 January 2019, Ghent, BE

Speech intelligibility prediction – The story continues

Richard Hendriks(a)
Delft University of Technology, The Netherlands

(a) Presenting

Developed about a century ago, speech intelligibility prediction has been dominated for a long time by the Articulation Index (AI), and, later on by its successor, the Speech Intelligibility Index. Although effective in some situations, these measures have difficulties to account for distortions introduced due to e.g., non-stationary noise and enhancement algorithms. Therefore, during the last decade, we have seen a significant increase in the interest on speech intelligibility measures, which led to the introduction of many new measures. Although the general trend is that these measures lead to improved prediction in certain specific situations, they typically only work well for a very narrow set of conditions and are often developed based on experience, hampering generalizations to new environments and processing conditions. Moreover, while the AI was quite close to a measure on the maximum information rate that may be transmitted from a talker to a listener, the majority of the more recent intelligibility metrics have lost this information theoretical interpretation.

In this presentation we use information theoretical concepts to model the transfer of information from the talker to the listener. Based on a rather simple model of communication, we present a new intelligibility metric that expresses the intelligibility as the estimate of the information shared between a talker and a listener in bits/sec. This measure, named Speech Intelligibility in Bits (SIIB), shows a very high correlation with speech intelligibility under a wide range of processing conditions.

Last modified 2019-01-08 16:51:41