This disclosure describes techniques to correct errors in automatic speech recognition, e.g., as performed to recognize spoken queries from a user to a virtual assistant or other application. A machine learning model detects potentially misrecognized n-grams within transcribed text which are then underlined in a user interface. A user can tap on the underlined n-gram, or another portion of the transcribed text to activate a dropdown menu that presents alternatives to the transcribed text. The alternatives can be based on speech hypothesis scores. To correct the error in transcribed text, the user picks an alternative from the dropdown menu, or, in the absence of a suitable alternative, types in the correction. With user permission, the error and corresponding correction are used as training data to improve model performance.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.