Speaker Verification Neural Network D-Vector Enrollment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speaker verification systems require retraining when enrolling new users, which is inefficient and may not be suitable for mobile devices with limited memory, and they struggle to handle a large number of training speakers without increasing the model size.
Innovation Solution
A speaker verification model based on a neural network is trained using utterances from multiple speakers, allowing new users to enroll without retraining, with a compact model size suitable for mobile devices, using the last hidden layer's output to generate a d-vector for verification, and combining it with i-vector techniques for improved performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing speaker verification systems retrain the model for each new user, then verification accuracy is improved, but system complexity and processing time increase
Solution Approach 1:
The system segments the speaker verification process into two independent parts: (1) a pre-trained neural network model that captures general speech characteristics, and (2) a user-specific reference vector extracted during enrollment. This segmentation allows the model to be trained once on diverse speakers while enabling rapid enrollment of new users without retraining, resolving the contradiction between accuracy and complexity.
Solution Approach 2:
The neural network model is pre-trained in advance using utterances from a large number of speakers before deployment. This preliminary action creates a robust foundation that can generalize to new speakers, eliminating the need for retraining when enrolling new users while maintaining high verification accuracy.
2Adaptability or versatility
If the speaker verification model is trained with a large number of training speakers, then generalization capability is improved, but model size increases
Solution Approach 1:
The system extracts only the essential user-specific information from the neural network's hidden layer to create a compact reference vector during enrollment. This extraction approach allows the model to leverage the generalization capability of a large-scale pre-trained network while storing minimal user-specific data, resolving the contradiction between generalization and model size.
Solution Approach 2:
Instead of storing complete model parameters for each user, the system creates a simplified copy or representation (reference vector) of user-specific characteristics by passing enrollment utterances through the pre-trained network. This copying mechanism enables support for many users without proportionally increasing storage requirements.
3Quantity of substance
If the speaker verification model is optimized for mobile devices with limited memory, then memory efficiency is improved, but processing capability is reduced
Solution Approach 1:
The system segments processing into two phases: (1) intensive pre-training performed offline on powerful servers, and (2) lightweight inference performed on mobile devices using only the pre-trained model and compact reference vectors. This segmentation enables mobile devices to operate with limited memory while maintaining verification capability.
Solution Approach 2:
The pre-trained neural network acts as an intermediary that performs the computationally intensive feature extraction during offline training. During mobile device operation, this intermediary has already captured complex speech patterns, allowing the device to perform only simple vector comparisons with minimal processing power and memory requirements.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for inputting speech data that corresponds to a particular utterance to a neural network; determining an evaluation vector based on output at a hidden layer of the neural network; comparing the evaluation vector with a reference vector that corresponds to a past utterance of a particular speaker; and based on comparing the evaluation vector and the reference vector, determining whether the particular utterance was likely spoken by the particular speaker.