Memory DNN Bottleneck Feature Extraction for Voiceprint Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voiceprint recognition systems lack comprehensive feature extraction, leading to inadequate accuracy in speaker identification due to insufficient representation of acoustic features.
Innovation Solution
A method involving a deep neural network with a memory function, including temporal recurrent layers and a bottleneck layer, is used to extract bottleneck features that are concatenated with spectral features to form more comprehensive acoustic features, improving the extraction of identity authentication vectors for enhanced speaker recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional spectral feature extraction is used for voiceprint recognition, then the system is simple and fast, but the recognition accuracy is insufficient due to inadequate feature representation
Solution Approach 1:
The patent combines traditional spectral features (MFCC with first and second order differences) with bottleneck features from a memory DNN to form comprehensive acoustic features. This merging of multiple feature sources resolves the contradiction by achieving both high recognition accuracy through comprehensive feature representation and maintaining system simplicity through the structured combination approach
Solution Approach 2:
The patent introduces a bottleneck layer with reduced dimensions (64 dimensions) between the temporal recurrent layer and subsequent layers. This dimensional transformation creates a compressed feature representation that captures essential speaker characteristics while reducing complexity, thereby improving recognition accuracy without proportionally increasing system complexity
2Measurement precision
If comprehensive acoustic features are extracted using memory DNN with bottleneck layer, then the voiceprint recognition accuracy improves, but the computational complexity and processing time increase
Solution Approach 1:
The patent extracts only the essential bottleneck features from the memory DNN's bottleneck layer, rather than processing all features through the entire network. This extraction of key features at the bottleneck layer reduces processing time while maintaining the accuracy benefits of comprehensive feature representation
Solution Approach 2:
The patent performs preliminary feature extraction by computing spectral features and feeding them into the memory DNN to obtain bottleneck features before combining with spectral features. This preliminary processing organizes the data in advance, making the subsequent combination and recognition processes more efficient
3Loss of information
If only spectral features are used for acoustic feature formation, then the feature extraction process is simple, but the representation of timing-dependent prosodic features is insufficient
Solution Approach 1:
The patent segments the feature extraction process into distinct components: spectral feature extraction, bottleneck feature extraction from memory DNN, and combination of both. The memory DNN is specifically designed with temporal recurrent layers to segment and process timing-dependent information separately, thereby capturing prosodic features without requiring complete structural complexity throughout the entire system
Solution Approach 2:
The bottleneck layer acts as an intermediary between the temporal recurrent layer and the final acoustic feature formation. It transforms the temporal features into a compressed representation that specifically captures timing-dependent prosodic information, serving as a mediator that adds this critical information without requiring the full complexity of the temporal recurrent network in the final feature set
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
Implementations of the present specification provide a voiceprint recognition method and device. The method includes: extracting a first spectral feature from speaker audio; inputting the speaker audio to a memory deep neural network (DNN), and extracting a bottleneck feature from a bottleneck layer of the memory DNN, where the memory DNN includes at least one temporal recurrent layer and the bottleneck layer, an output of the at least one temporal recurrent layer is connected to the bottleneck layer; forming an acoustic feature of the speaker audio based on the first spectral feature and the bottleneck feature; extracting an identity authentication vector corresponding to the speaker audio based on the acoustic feature; and performing speaker recognition by using a classification model and based on an identity authentication vector (i-vector).