Abstract
Speech biasing refers to specifying hints to a natural language understanding (NLU) algorithm or otherwise configuring the speech recognizer to interpret an uttered phrase to match the user’s intent. Speech biasing is generally on a per-user basis and does not use other (or similar) user behavior to correct speech recognition errors or predict new phrases. This disclosure describes techniques to augment the set of candidate speech biasing phrases for recognizing the speech of a particular user with phrases fetched from similar users. A collection of users similar to the particular user is identified and corresponding speech biasing phrases are fetched. The phrases are ranked and filtered, and the resulting phrases are used as additional, candidate biasing phrases for the particular user. The techniques can run in real time and can incorporate prior queries from a user to identify similar users in a latency-sensitive manner.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Weisz, Ágoston and Salama, Khalid, "Speech Recognition Customization Based on User Embeddings", Technical Disclosure Commons, (February 23, 2023)
https://www.tdcommons.org/dpubs_series/5693