There are probably some specific measures concerning language speech recognition in smart phones. However, I would start with the known measures and adjusting them considering your specific semantics. Assuming it is insufficient to assess classification results using only Precision and Recall, you could take a look on how my heuristics for assessing classification results work. These are a classes-similarity measure, cost-based measure, classes-number measure in my thesis about emotion recognition -- Thesis Opinion Mining and Lexical Affect Sensing

I found a similar question here:

https://www.researchgate.net/post/What_are_the_performance_measures_in_Speech_recognition