Speaker Recognition UBM Training via GMM Component Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speaker recognition systems using Gaussian Mixture Models (GMMs) are resource-intensive, requiring significant processor power and memory, which limits their efficiency and computation speed, necessitating a solution to reduce resource usage and improve processing times.
Innovation Solution
The implementation of an efficient universal background model (UBM) training method that involves receiving audio inputs, extracting identifying features, generating feature vectors, and performing optimized training sequence computations based on a GMM with components defined by covariance matrices, mean vectors, and weight vectors, while utilizing a resource optimizer to reduce redundant calculations and memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Gaussian Mixture Models (GMMs) are used for speaker recognition, then recognition accuracy is improved, but processor power and memory requirements increase significantly
Solution Approach 1:
The patent segments the GMM training process into distinct phases: UBM training phase and speaker adaptation phase. During speaker adaptation, only a subset of GMM components (those with highest posterior probabilities) are updated rather than all components, significantly reducing computational load while maintaining recognition accuracy
Solution Approach 2:
The patent applies partial action by updating only the most relevant GMM components during speaker adaptation. Specifically, components with posterior probability above a threshold are selected for updates, avoiding redundant computations on less relevant components while preserving the essential recognition functionality
2Reliability
If traditional GMM training methods are used, then model accuracy is maintained, but computation time increases
Solution Approach 1:
The patent performs UBM training in advance using all available data to establish a comprehensive baseline model. This preliminary action allows the system to have a well-trained GMM before actual speaker recognition tasks, enabling faster real-time processing during the adaptation phase without compromising final model accuracy
Solution Approach 2:
The training process is divided into offline UBM training and online speaker adaptation. The computationally intensive UBM training is performed once offline, while speaker-specific adaptations are performed quickly online by updating only relevant components, thus reducing real-time computation time
3Reliability
If comprehensive GMM component updates are performed, then model robustness is improved, but memory usage increases
Solution Approach 1:
The patent extracts and updates only the most relevant GMM components during speaker adaptation by selecting components with highest posterior probabilities. This extraction approach maintains model robustness by focusing on the most informative components while reducing memory usage by avoiding storage and processing of redundant component data
Data Source
AI summary
A system and method for efficient universal background model (UBM) training for speaker recognition, including: receiving an audio input, divisible into a plurality of audio frames, wherein at least a first audio frame of the plurality of audio frames includes an audio sample having a length above a first threshold extracting at least one identifying feature from the first audio frame and generating a feature vector based on the at least one identifying feature; generating an optimized training sequence computation based on the feature vector and a Gaussian Mixture Model (GMM), wherein the GMM is associated with a plurality of components, wherein each of the plurality of components is defined by a covariance matrix, a mean vector, and a weight vector; and updating any of the associated components of the GMM based on the generated optimized training sequence computation.