Training Data-Driven Speech Intelligibility Predictors on Heterogeneous Listening Test Data

Prediction of Speech Intelligibility (SI) is a topic interest for most speech processing applications, where intelligibility any importance, e.g., coding, transmission and enhancement. Traditionally, SI predictors have been based on signal methods heuristics, but more recently, an increasing number data-driven SI-predictors proposed. Data-driven prediction requires large quantities labelled data, ideally from many listening tests. Listening tests differ in factors such as vocabulary, talker, listener’s task, etc. collectively referred to the paradigm. A naïve strategy training directly stimuli, pooled different tests, futile because exact map stimulus determined, not only by stimulus, also trained this way become specialized paradigms data erroneously attributing all paradigm influences stimulus. The problem fundamental persists even idealized situation abundant. We propose that independent paradigms, underlying data. proposed concatenate SI-predictor layer trainable dataset-specific mapping functions, each corresponding single These functions are jointly with serve efficiently approximate psychometric implied prevent predictor specializing these during training. present novel architecture incorporates convolutional network ESTOI back-end, train it strategy, compare range existing non-data-driven predictors. results higher performance overall increased robustness unseen paradigms.