Article "Optimizing Automatic Speech Recognition for Low-Proficient Non-Native Speakers" (9th)

Clarisa livia
16611022


This study conducted two experiments to evaluate methods for speech utterances and test utterances that will be used in CALL applications for low-educated Dutch L2 students. For speech selection with transcription responses in the language model, their best error rate is between 10.0% and 6.9% after optimizing the acoustic and language model.

For example, if it is unclear whether a segment or a word (short) is spoken or not, this can be ascertained in step two through more detailed analysis. This is very encouraging, especially if we remember that in our language learning applications we can be conservative, we are not quite sure of the results we can always ask the language learner to try again.


However, the level of seriousness will depend on the degree of discrepancy between the actual speech produced and which is recognized and indicated by the system: the greater the deviation the more serious the error. On the other hand, large deviations are less likely than small deviations. On the basis of such considerations, we can show the seriousness of both types of errors and therefore the costs to be paid for false rejection and false acceptance.