Using Smoothed Heteroscedastic Linear Discriminant Analysis in Large Vocabulary Continuous Speech Recognition System?

Abstract. In the state-of-the-art speech recognition systems, Het-eroscedastic Linear Discriminant Analysis (HLDA) is becoming popular technique allowing for feature decorrelation and dimensionality reduc-tion. However, HLDA relies on statistics, which may not be reliably esti-mated when only limited amount of training data is available. Recently, Smoothed HLDA (SHLDA) was proposed as a robust modi¯cation of HLDA. Previously, SHLDA was successfully used for feature combina-tion in small vocabulary recognition experiments [1]. In this work, we verify that SHLDA can be advantageously used also for Large Vocabu-lary Continuous Speech Recognition.