Information transmission analysis for continuous speech features

Acoustic features are important for the study of human speech perception and the development of computational speech classification algorithms. These areas can benefit from a technique that accurately quantifies the information content of an individual feature, as well as the degree to which this information is used by a listener or algorithm. Feature information transmission analysis (FITA) was developed to do this for categorical features, but complications arise when applying it to continuous features. Absolute information measures are bounded from above by values substantially lower than their theoretical maxima, the precision with which a feature is produced is ignored and a tedious manual step is introduced into the analysis process. This article presents an alternative approach (the continuous FITA) that addresses these complications effectively by representing continuous features in a more natural way. It is shown that this approach is well-suited to continuous feature information analysis and, furthermore, can be used to estimate redundancy in multiple features and information transmitted by combinations of features. The continuous FITA can quantify information in extracted features before an identification experiment has been conducted. It can aid in feature selection for computational speech classification systems, measure feature utilization by humans as well as computational algorithms and facilitate the investigation of the effect of noise or signal processing on feature information.