To evaluate whether the adapted acoustic model improves accuracy, compare the speech-to-text results with a control run performed with the standard acoustic model. For a fair comparison, you must run the evaluation on audio that was not included in the adaptation set.
You can also run the test on the adaptation data as a sanity check on the adaptation process. In this case, there should be a significant improvement in the performance, because the model has been adapted to the exact data that it is now being tested on. If the new model is not significantly better that the standard model, the adaptation process might not have functioned adequately.
For more information about evaluating performance, see Score Speech Recognition.
|