Evaluate a Custom Language Model

After training a custom language model you can evaluate its performance. To do this you need an audio file containing some speech, and an accurate transcript.

To evaluate your custom language model, use the action ScoreCustomSpeechLanguageModel. You supply the speech-to-text settings that you want to use, and Media Server runs speech-to-text using your language model. When this is complete, Media Server compares the output to the transcript and outputs some statistics - the F-measure, precision, and recall.

If the custom language model does not provide the accuracy you require, you might re-train the model with more text. You can evaluate the model after each change, to see whether the accuracy has improved. When you are satisfied with the accuracy, you can stop training.