DNN Dimensionality Reduction for Speaker Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speaker recognition systems using i-vectors are time-consuming at both training and testing times, and are not well-suited for variable-duration utterances, with existing post-processing techniques failing under high speaker and channel variability conditions.
Innovation Solution
A deep neural network (DNN) is employed to reduce the dimensionality of first-order Gaussian statistics, using fully connected layers to normalize and process audio features, and a dropout technique to handle variable utterance durations, along with a triplet architecture for channel-independent voiceprint generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional i-vector based speaker recognition systems are used, then speaker recognition functionality is achieved, but computation time is excessive and the system is not suitable for real-time applications
Solution Approach 1:
The patent changes the computational parameters by replacing the conventional i-vector extraction and PLDA scoring pipeline with a deep neural network that directly processes Baum-Welch statistics. The DNN learns optimal feature transformations and speaker discrimination in an end-to-end manner, fundamentally changing how the recognition task is performed to achieve both speed and accuracy
Solution Approach 2:
The patent substitutes the mechanical multi-stage processing system (UBM training, i-vector extraction, channel compensation, PLDA scoring) with a neural network-based system. The DNN replaces multiple sequential computational stages with a unified learned model that achieves the same functionality more efficiently
2Reliability
If post-processing techniques like whitening, LDA, and PLDA are applied to compensate for channel variability, then channel robustness is improved, but the system fails under high speaker and channel variability conditions
Solution Approach 1:
The patent introduces dynamic adaptability through the deep neural network that can learn and adapt to various speaker and channel conditions during training. The DNN's learned parameters automatically adjust to handle high variability scenarios, making the system dynamically adaptable rather than relying on fixed post-processing transformations
Solution Approach 2:
The patent performs preliminary learning of speaker and channel variability patterns during the DNN training phase. By pre-training on diverse data with high speaker and channel variability, the network learns robust representations before actual recognition, preparing it to handle such conditions without requiring adaptive post-processing
3Loss of information
If i-vectors are computed using total variability procedure, then speaker and channel information is captured, but the training and testing processes are time-consuming
Solution Approach 1:
The patent merges the separate functions of i-vector extraction, channel compensation, and speaker verification into a single deep neural network. This consolidation eliminates multiple processing stages and their associated computational overhead, achieving the same information preservation goals faster by combining operations that were previously performed sequentially
Data Source
AI summary
In a speaker recognition apparatus, audio features are extracted from a received recognition speech signal, and first order Gaussian mixture model (GMM) statistics are generated therefrom based on a universal background model that includes a plurality of speaker models. The first order GMM statistics are normalized with regard to a duration of the received speech signal. The deep neural network reduces a dimensionality of the normalized first order GMM statistics, and outputs a voiceprint corresponding to the recognition speech signal.