Learn More
Theoretically, an improvement in a language model occurs as the size of the n-grams increases from 3 to 5 or higher. As the n-gram size increases, the number of parameters and calculations, and the storage requirement increase very rapidly if we attempt to store all possible combinations of n-grams. To avoid these problems, the reduced n-grams' approach(More)
Queens Univer s ity Belf a s t, Nor ther n I r ela nd Abstrac t Er r or s ca n occur a t ever y level of a dia logue, f r om the r ecognition of wha t wor ds wer e s poken to the under s ta nding of the intentions behind the wor ds. Our a ppr oa ch to er r or-ha ndling a s s umes tha t er r or s ca nnot be a voided in s poken dia logue a nd tha t it is mor(More)
This paper investigates the effect of modeling subband correlation for noisy speech recognition. Subband feature streams are assumed to be independent in many subband-based speech recognition systems. However, speech recognition experimental results suggest this assumption is unrealistic. In this paper, a method is proposed to incorporate correlation into(More)
Advanced spoken dialogue systems incorporate functionalities such as mixed-initiative and cross-domain dialogues. In this paper an object-based approach to cross domain dialogue modelling is described in which service agents representing primary transaction types and support agents representing tasks such as eliciting payment details are selected as(More)
In this article we describe how Java can be used to implement an object-based, cross-domain, mixed initiative spoken dialogue manager (DM). We describe how dialogue that crosses between several business domains can be modelled as an inheriting and collaborating suite of objects suitable for implementation in Java. We describe the main features of the Java(More)
This paper proposes a new statistical approach for recognizing speech with unknown, burst noise during the utterance. The new model characterizes the partially and randomly corrupted observations based on the union of random events. Currently we have tested the new model using a speaker-independent telephony digit database, corrupted by various types of(More)
Most current speech recognition systems are built upon a single type of model, e.g. an HMM or certain type of segment based model, and furthermore typically employs only one type of acoustic feature e.g. MFCCs and their variants. This entails that the system may not be robust should the modeling assumptions be violated. Recent research efforts have(More)
The Zipf curves of log of frequency against log of rank for a large English corpus of 500 million word tokens, 689,000 word types and for a large Spanish corpus of 16 million word tokens, 139,000 word types are shown to have the usual slope close to –1 for rank less than 5,000, but then for a higher rank they turn to give a slope close to –2. This is(More)