Multilingual acoustic model training based on phoneme set conversion

For multilingual automatic speech recognition (ASR), acoustic models are often trained on multiple language datasets. The ASR of a target language can be improved by using auxiliary language data. Generally, universal phoneme sets are used in order to eliminate the differences between phoneme sets. However, this approach is inefficient for languages whose phoneme sets are similar. We propose a more efficient training method based on phoneme set conversion, in which phonemes from the auxiliary language are converted into those of the target language. In addition, we examined the construction of context trees.