Individualized Hotword Detection Models Using Candidate Acoustic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice command systems struggle to accurately detect hotwords due to variations in pronunciation among users, leading to missed detections, and require a large number of enrollment utterances for generating effective hotword detection models.
Innovation Solution
A method for generating individualized hotword detection models by obtaining enrollment acoustic data from a user, selecting similar utterances from other users based on similarity scores, and training a neural network using these selected utterances to improve detection accuracy without needing thousands of enrollment utterances from the user.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the system generates individualized hotword detection models using only enrollment utterances from the user, then the model accuracy for detecting the user's hotword is improved, but the number of required enrollment utterances increases significantly
Solution Approach 1:
The patent introduces candidate utterances from other users as an intermediary data source. Instead of requiring only the target user's enrollment utterances, the system uses acoustic data from multiple other users speaking the same hotword to supplement and expand the training data, thereby reducing the quantity of enrollment utterances needed from the target user while maintaining detection accuracy.
Solution Approach 2:
The patent makes the hotword detection model universally applicable across multiple users by incorporating acoustic data from various users into the training process. The model learns to recognize hotwords spoken by different users, enabling it to serve multiple users with their respective pronunciation variations without requiring separate models for each user.
2Device complexity
If the system uses a generic hotword detection model, then the device complexity is reduced, but the reliability of hotword detection decreases due to pronunciation variations among users
Solution Approach 1:
The patent applies local quality by customizing the hotword detection model to match individual user's pronunciation characteristics. Instead of using a single generic model, the system generates a personalized model for each user based on their enrollment utterances and the selected candidate utterances, thereby improving detection reliability for each user's specific pronunciation variations.
Solution Approach 2:
The patent changes the parameters of the detection model by adjusting it to fit each user's acoustic characteristics. The model parameters are updated based on the user's enrollment data and the selected candidate utterances, transforming a generic model into a user-specific model that adapts to individual pronunciation patterns.
3Measurement precision
If the system requires users to provide thousands of enrollment utterances, then the detection model accuracy is improved, but the ease of operation deteriorates due to extensive user input requirements
Solution Approach 1:
The patent uses candidate utterances from other users as an intermediary to bridge the data gap. This intermediary source provides additional training data without requiring the target user to provide thousands of utterances, thereby maintaining detection model accuracy while significantly improving ease of operation by reducing the user input burden.
Solution Approach 2:
The patent copies acoustic data from other users who have spoken the hotword before. These copied utterances serve as proxy data that supplements the limited enrollment data from the target user, enabling the system to achieve high detection accuracy without requiring extensive new input from the user.
4Adaptability or versatility
If the system processes all candidate acoustic data without selection, then the detection model comprehensiveness is improved, but the computational resources required increase significantly
Solution Approach 1:
The patent applies partial action by selecting only a subset of candidate utterances rather than processing all available data. The system uses similarity scores to identify and process only the most relevant candidate utterances that are similar to the user's enrollment utterances, thereby reducing computational resources while maintaining sufficient model comprehensiveness.
Solution Approach 2:
The patent extracts and processes only the essential candidate utterances that are most similar to the user's enrollment data. By filtering and extracting only the relevant portion of candidate data based on similarity thresholds, the system reduces the computational load while preserving the necessary comprehensiveness for accurate detection.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for presenting notifications in an enterprise system. In one aspect, a method include actions of obtaining enrollment acoustic data representing an enrollment utterance spoken by a user, obtaining a set of candidate acoustic data representing utterances spoken by other users, determining, for each candidate acoustic data of the set of candidate acoustic data, a similarity score that represents a similarity between the enrollment acoustic data and the candidate acoustic data, selecting a subset of candidate acoustic data from the set of candidate acoustic data based at least on the similarity scores, generating a detection model based on the subset of candidate acoustic data, and providing the detection model for use in detecting an utterance spoken by the user.