Frank Wessel

Statistical Language Modeling Using Leaving-One-Out

The need for a stochastic language model in speech recognition arises from Bayes’ decision rule for minimum error rate (Bahl et al., 1983). The word sequence w1 ... w N to be recognized from the sequence of acoustic observations x 1 ... x T is determined as that word sequence w 1 ... w N for which the posterior probability Pr(w 1 ... w N |x 1 ... x T ) attains its maximum. This rule can be rewritten in the form:

Robust dialogue-state dependent language modeling using leaving-one-out

\mathop {\arg \,\max }\limits_{{\omega _1} \cdot \cdot \cdot {\omega _N}} \{ \Pr ({\omega _1} \cdot \cdot \cdot {\omega _N}) \cdot \Pr ({x_1} \cdot \cdot \cdot {x_T}\left| {{\omega _1} \cdot \cdot \cdot {\omega _N}} \right.)\} ,

Confidence measures for large vocabulary continuous speech recognition

, where Pr(x 1 ... x T |w 1 ... w N ) is the conditional probability of, given the word sequence w 1 ... w N , observing the sequence of acoustic measurements x 1 ... x T and where Pr(w 1 ... w N ) is the prior probability of producing the word sequence w 1 ... w N .

Using word probabilities as confidence measures

We present an overview of the RWTH Aachen large vocabulary continuous speech recognizer. The recognizer is based on continuous density hidden Markov models and a time-synchronous left-to-right beam search strategy. Experimental results on the ARPA Wall Street Journal (WSJ) corpus verify the effects of several system components, namely linear discriminant analysis, vocal tract normalization, pronunciation lexicon and cross-word triphones, on the recognition performance.

Unsupervised training of acoustic models for large vocabulary continuous speech recognition

The use of dialogue-state dependent language models in automatic inquiry systems can improve speech recognition and understanding if a reasonable prediction of the dialogue state is feasible. In this paper, the dialogue state is defined as the set of parameters which are contained in the system prompt. For each dialogue state a separate language model is constructed. In order to obtain robust language models despite the small amount of training data we propose to interpolate all of the dialogue-state dependent language models linearly for each dialogue state and to train the large number of resulting interpolation weights with the EM-algorithm in combination with leaving-one-out. We present experimental results on a small Dutch corpus which has been recorded in the Netherlands with a train timetable information system and show that the perplexity and the word error rate can be reduced significantly.

Explicit word error minimization using word hypothesis posterior probabilities

This paper gives an overview of an architecture and search organization for large vocabulary, continuous speech recognition (LVCSR at RWTH). In the first part of the paper, we describe the principle and architecture of a LVCSR system. In particular, the isssues of modeling and search for phoneme based recognition are discussed. In the second part, we review the word conditioned lexical tree search algorithm from the viewpoint of how the search space is organized. Further, we extend this method to produce high quality word graphs. Finally, we present some recognition results on the ARPA North American Business (NAB’94) task for a 64 000-word vocabulary (American English, continuous speech, speaker independent).