Cluster-Specific Speech Model for Mobile Voice Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies face challenges in accurately transcribing user utterances due to variations in acoustic characteristics such as gender, accent, pitch, and background noise, leading to suboptimal recognition performance.
Innovation Solution
The approach involves representing acoustic characteristics as vectors in a vector space, segmenting the space into clusters, and training a neural network-based speech model for each cluster. A new user's utterance is mapped to a cluster, and the corresponding speech model is selected and provided for recognition, allowing for more accurate and compact speech recognition models that can be loaded on mobile devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single general speech model is used for all users, then the device complexity is reduced, but the speech recognition accuracy deteriorates due to variations in acoustic characteristics
Solution Approach 1:
The patent segments the vector space into multiple clusters based on acoustic characteristics, with each cluster having its own dedicated speech model. This segmentation allows the system to achieve high recognition accuracy for diverse users while keeping individual models relatively simple, as each model only needs to handle a specific acoustic profile rather than all possible variations.
Solution Approach 2:
The patent applies local quality by training each speech model specifically for its corresponding cluster's acoustic characteristics. Each model is optimized for local acoustic patterns (e.g., male voices, female voices, specific accents) rather than being a general-purpose model, thereby achieving high accuracy for its target group while maintaining manageable complexity.
2Measurement precision
If multiple speech models are trained for different acoustic characteristics, then the speech recognition accuracy is improved, but the device complexity and memory requirements increase
Solution Approach 1:
By segmenting users into acoustic clusters and assigning dedicated models to each cluster, the system achieves high recognition accuracy for diverse users while keeping individual models relatively compact. The segmentation allows efficient memory management where only the necessary number of specialized models need to be stored rather than one large general model.
Solution Approach 2:
The patent implements a universal cluster selection mechanism that can route any user's speech to the appropriate specialized model. The system maintains a set of specialized models for different acoustic characteristics but uses a universal clustering algorithm to determine which model to apply, achieving both specialization and efficiency.
3Measurement precision
If speech models are optimized for specific acoustic characteristics, then the recognition accuracy for those characteristics is improved, but the adaptability to new acoustic patterns deteriorates
Solution Approach 1:
The patent performs preliminary clustering of the vector space into multiple acoustic clusters before receiving user speech. This preliminary segmentation creates a framework that can accommodate new users by assigning them to the most appropriate existing cluster based on their acoustic characteristics, thus maintaining adaptability while using specialized models.
Solution Approach 2:
The system dynamically selects which speech model to use based on the acoustic characteristics of the incoming speech. Rather than being static, the system adapts by computing the user's acoustic profile and routing to the most suitable cluster-model pair, allowing the system to handle new acoustic patterns while leveraging pre-trained specialized models.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for receiving data representing acoustic characteristics of a user's voice; selecting a cluster for the data from among a plurality of clusters, where each cluster includes a plurality of vectors, and where each cluster is associated with a speech model trained by a neural network using at least one or more vectors of the plurality of vectors in the respective cluster; and in response to receiving one or more utterances of the user, providing the speech model associated with the cluster for transcribing the one or more utterances.