8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


Impact of Word Graph Density on the Quality of Posterior Probability Based Confidence Measures

Tibor Fabian, Robert Lieb, Gunther Ruske, Matthias Thomae

Technical University of Munich, Germany

Our new experimental results, presented in this paper, clearly prove the dependence between word graph density and the quality of two different confidence measures. Both confidence measures are based on the computation of the posterior probabilities of the hypothesized words and apply the time alignment information of the word graph for confidence score accumulation. We show that the quality of the confidence scores of both confidence measures significantly increases for higher word graph densities. The analyses were carried out on two different German spontaneous speech corpora: on the Verbmobil evaluation corpus [1] and on the NaDia corpus. We achieved a relative reduction of the confidence error rate by up to 41.4%, compared to the baseline confidence error rate. The results lead us to propose to perform the confidence score calculation - based on posterior probability accumulation - on higher word graph densities in order to get the best results.

Full Paper

Bibliographic reference.  Fabian, Tibor / Lieb, Robert / Ruske, Gunther / Thomae, Matthias (2003): "Impact of word graph density on the quality of posterior probability based confidence measures", In EUROSPEECH-2003, 917-920.