Triplet Neural Network Voiceprints for Speaker Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speaker recognition technologies based on universal background models and Gaussian mixture modeling discard useful information by relying on handcrafted features, making it difficult to model complex structures in the feature space.
Innovation Solution
A deep neural network with a triplet network architecture is used to train a front-end feature extractor for speaker verification and identification, utilizing a batch process to refine connection weights through a loss function based on similarity distances and margins.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If handcrafted features (MFCCs) and Gaussian mixture modeling are used, then the system is computationally efficient and easier to implement, but useful information is discarded and complex structures in feature space cannot be modeled
Solution Approach 1:
The patent replaces traditional handcrafted feature extraction and Gaussian mixture modeling with a deep neural network system. The DNN automatically learns and extracts features from raw speech signals, substituting the manual feature engineering process (MFCCs) and probabilistic modeling (GMM) with a data-driven approach that preserves more information while handling complex structures in the feature space.
2Reliability
If deep neural network with triplet network architecture is used, then speaker recognition performance is improved, but computational complexity and training requirements increase
Solution Approach 1:
The patent divides the deep neural network into three separate feed-forward neural networks (triplet architecture), each processing different input types (enrollment speech, verification speech, and cohort speech). This segmentation allows the system to handle different computational tasks in parallel, improving speaker recognition performance while organizing the complexity into manageable, modular components with shared connection weights.
Data Source
AI summary
The present invention is directed to a deep neural network (DNN) having a triplet network architecture, which is suitable to perform speaker recognition. In particular, the DNN includes three feed-forward neural networks, which are trained according to a batch process utilizing a cohort set of negative training samples. After each batch of training samples is processed, the DNN may be trained according to a loss function, e.g., utilizing a cosine measure of similarity between respective samples, along with positive and negative margins, to provide a robust representation of voiceprints.