Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Modelling Phonetic Context Using Head-Body-Tail Models for Connected Digit Recognition

Janienke Sturm, Eric Sanders

A2RT, Dept. Language and Speech, University of Nijmegen, The Netherlands

Both whole word modelling and context modelling have proven to improve recognition performance for connected digit strings. In this paper we will show that word boundary variation can be effectively modelled by applying the Head-Body-Tail (HBT) method as proposed by Chou et al in [1] and also applied by Gandhi in [2]. Each digit is split into three parts, representing the beginning, middle and end of a word. The middle part - the body - is assumed to be context-independent, whereas the first part - the head - and the last part - the tail - incorporate information about the preceding or subsequent digit. The results we obtained with HBT-modelling are compared with results obtained with whole-word models (WWMs) [3] and with the results obtained with HBT-models reported in [2]. It is shown that using HBT models a relative improvement over contextindependent WWMs of 28% on string level can be reached.


Full Paper

Bibliographic reference.  Sturm, Janienke / Sanders, Eric (2000): "Modelling phonetic context using head-body-tail models for connected digit recognition", In ICSLP-2000, vol.1, 429-432.