When a new voice feature is to be launched on a device with a voice interface, e.g., a digital assistant application, the natural language understanding (NLU) model is built using training data for the new feature. Speech biasing models are typically added to improve recognition accuracy for queries that are specific to the feature or contain non-common words. Such biasing models are often built using traffic logs, collected with user permission, after the initial release of the feature. However, this approach may not provide high speech recognition quality during product testing and initial launch. This disclosure describes techniques to improve the ASR quality of a new feature from the time of initial release and without relying on traffic logs. To that end, speech biasing models are built using grammar training phrases.

Creative Commons License

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