Analysing confusion matrices through the computation of Information Transfer Rates : An illustration based on a study of the perception of nasal consonants and vowels in channel-vocoded speech.

Olivier Crouzet (LLING, UMR6310 Université de Nantes / CNRS)
09 décembre 2016, 14h0015h30

The analysis of human speech perception data often relies on the analysis of overall performance but also on the interpretation of « qualitative » errors based on the exploration of « confusion matrices ». Such confusions are also frequently used in « machine » or « statistical » classification experiments (using either supervised or unsupervised methods, e.g. LDA, logistic regression, clustering…). Within the framework of a project investigating the perception of nasal consonants and vowels in « channel-vocoded » speech and cochlear implanted deaf listeners, I have developed an R package that is devoted to computing « Information Transfer Rate » analyses as described in Miller & Nicely (1955) : a mathematical tool that provides quantitative measurements of qualitative classification errors. Though this approach was introduced as early as the mid-50s by Miller & Nicely (1955) based on Shannon’s (1948) Information Theory and later extended by Wang & Bilger (1973), these tools have recently been reintroduced in the analysis of speech classification tasks (see Christiansen & Greenberg, 2012). I have since started developing an R package (iteR) that is devoted to help users analyse and manipulate confusion matrices within this framework. This package will also be compared with David van Leuwen’s sinfa.R script (https://github.com/davidavdav/sinfa) which is itself based on Wang & Bilger (1973)’s work. Both tools are complementary and iteR may later integrate sinfa.R procedures.

Prochains événements

Voir la liste d'événements