Predictability of the effects of phoneme merging on speech recognition performance by quantifying phoneme relations

Abstract:

To investigate whether the impact of phoneme merging on recognition rate can be predicted, different measures to quantify the relationship between two phonemes a and b have been compared: (1) the functional load of their opposition, (2) the bigram type preservation, (3) their information radius, (4) their distance within an information gain tree induced from a distinctive feature matrix, and (5) the symmetric Kullback-Leibler divergence. For each of 26 different phoneme pairs we trained a speech recognition system where the phoneme pair was merged. We then compaired the new accuracy rates and the measures to find out if there was any correlation. The results did not always meet our expectations and raised further questions.


Year: 2013
In session: Sprach- und Sprechererkennung
Pages: 247 to 253