Speaker Model Creation Using Intermediate Neural Network Layer Extraction
Solution Overview
Problem
Conventional speaker recognition using neural networks requires extensive input speech for registration, leading to increased calculation complexity and inefficiency, as well as the need for multiple neural networks and additional noise files, which can result in unnatural speech and reduced registration efficiency.
Innovation Solution
A system that utilizes outputs from a predetermined intermediate layer in a neural network as a feature for speaker recognition, changing network parameters from the input layer to the intermediate layer based on multiple patterns to extend input speech features, allowing for efficient speaker registration with less data and using a single neural network for keyword detection and recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional neural network speaker recognition is used, then accurate speaker recognition can be achieved, but extensive input speech is required for registration processing
Solution Approach 1:
The patent extracts speaker recognition features from an intermediate layer of the neural network during keyword detection, rather than requiring separate extensive speech samples. This extraction approach obtains speaker features as a byproduct of normal operation, significantly reducing the amount of speech data needed for registration while maintaining recognition accuracy.
Solution Approach 2:
The neural network is designed to perform multiple functions: keyword detection and speaker recognition simultaneously. By utilizing the same network for both purposes and extracting speaker features from an intermediate layer, the system eliminates the need for separate registration processing and reduces data requirements.
2Reliability
If conventional speaker recognition registration is performed, then speaker models can be created, but calculation complexity increases
Solution Approach 1:
The patent merges keyword detection and speaker recognition into a single unified process. By extracting speaker features from an intermediate layer of the keyword detection network, the system combines both functions, eliminating the need for separate registration processing and reducing overall calculation complexity.
Solution Approach 2:
The system performs speaker feature extraction automatically during normal keyword detection operations without requiring separate registration processing. The intermediate layer outputs are reused for speaker recognition, allowing the system to serve multiple purposes with the same computational resources.
3Measurement precision
If multiple neural networks are used for speaker recognition, then recognition accuracy can be improved, but device complexity increases
Solution Approach 1:
The patent makes a single neural network perform multiple functions by extracting speaker recognition features from an intermediate layer during keyword detection. This multi-functional approach eliminates the need for separate neural networks for different recognition tasks, reducing device complexity while maintaining recognition accuracy.
Solution Approach 2:
The neural network is segmented into functional layers, with an intermediate layer being extracted for speaker feature purposes. This segmentation allows the same network structure to serve dual purposes: keyword detection in the output layer and speaker recognition through intermediate layer extraction.
Data Source
AI summary
According to one embodiment, a system for creating a speaker model includes one or more processors. The processors change a part of network parameters from an input layer to a predetermined intermediate layer based on a plurality of patterns and inputs a piece of speech into each of neural networks so as to obtain a plurality of outputs from the intermediate layer. The part of network parameters of the each of the neural networks is changed based on one of the plurality of patterns. The processors create a speaker model with respect to one or more words detected from the speech based on the outputs.