Likelihood Probability Mismatch Analysis and Normalization in Multilingual Speech Applications

In this paper, with a multilingual speech recognition system, we exam the HMM likelihood scores among the different acoustic models and observe that there exist scoring mismatches. The mismatches might come from different recording environments in which the training data for each language were collected, or come from different acoustic modeling structures. This analysis helps us understand the gaps among the likelihood probabilities on these acoustic models. Based on the observation of the differences of likelihood probability scores from different languages, we study a simple frame based likelihood probability normalization method to balance the likelihood scores of multiple acoustic models in the recognition system. Experiments show that this normalization method is effective to compensate the likelihood probability biases that come from different training corpora and different acoustic structures.