EUROSPEECH 2001 Scandinavia
This paper describes an efficient method to predict words whose recognition rates are low. This method is based on the idea that the minimum value of the word pair recognition rates corresponds to the word recognition rate. The word pair recognition rates can be calculated by the measured distributions of phoneme log likelihood difference. The proposed method was evaluated by recognition experiments using about 3000 word pairs. The correlation coefficient between the predicted and the measured recognition rates was 0.87 when the phoneme lengths of both words of the word pair were equal. Furthermore, we also estimated a 95% confidence interval for the measured recognition rates, and the percentage of the predicted words that were contained in the confidence interval was 94.8%. The results showed the effectiveness of the proposed method for predicting the word pair recognition rates.
Bibliographic reference. Terashima, Ryuta / Hoshino, Hiroyuki / Wakita, Toshihiro (2001): "Prediction of low recognition rate words for isolated word recognition system", In EUROSPEECH-2001, 2095-2098.