Effective operation of a voice-activated virtual assistant requires accurate speech recognition. Manual determination of the accuracy of machine-generated speech transcriptions requires involvement of third parties to evaluate transcriptions of a user’s speech. Automated accuracy evaluation approaches that use machine-generated speech as input and determine quality of transcription have limited effectiveness since the machine-generated speech is a poor proxy for real-world user speech, e.g., volume of input, microphone and room characteristics, pronunciations, etc. This disclosure describes obtaining user confirmation of the transcription of a small subset user queries as performed by a virtual assistant or other applications that accept speech input. With user permission, the obtained data, e.g., the user verifying that the transcription was accurate or indicating that it was wrong, are used to rate and/or update the speech recognition technology, e.g., train speech recognition machine learning models.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Abdelaziz, Omar; Ogbonna, Justice; Weisz, Ágoston; Koerkamp, Ragnar Groot; Lu, Xinran; Krishnakumaran, Saisuresh; Baía, Tâmara; Vuskovic, Vladimir; and Yu, Tony, "Accuracy Evaluation Of Automated Speech Transcription Via User Feedback", Technical Disclosure Commons, (April 29, 2020)