Neural Network Speaker Verification Model for Mobile Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speaker verification systems face challenges in accurately identifying users based on voice characteristics, particularly in compact form factors like mobile devices, where memory and computational resources are limited, and require retraining for new users.
Innovation Solution
The implementation of a neural network-based speaker verification model and evaluation vector similarity model that uses labeled pairs of feature vectors to determine whether an utterance was spoken by a particular speaker, allowing for efficient enrollment and verification without retraining, and utilizing a compact speaker verification model suitable for client devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speaker verification systems are used, then speaker identification accuracy can be achieved, but the system requires significant memory and computational resources and needs retraining for new users
Solution Approach 1:
The system is divided into two independent neural networks: a speaker verification model that extracts features from speech data, and an evaluation vector similarity model that compares feature vectors. This segmentation allows each component to be optimized independently, reducing overall computational complexity while maintaining accuracy.
Solution Approach 2:
Instead of storing and processing raw speech data for comparison, the system extracts and stores compact feature vectors that represent the essential characteristics of speakers. These feature vectors are mathematical abstractions that capture speaker identity information in a compressed form, significantly reducing memory requirements.
2Adaptability or versatility
If traditional speaker verification systems are used, then speaker identification can be performed, but retraining is required for new users which increases system complexity
Solution Approach 1:
The speaker verification model is pre-trained on a diverse dataset of multiple speakers before deployment. This preliminary training enables the model to extract universal speaker-related features that work across different individuals. When new users are added, the system only needs to collect their speech samples and generate feature vectors, without requiring retraining of the core verification model.
Solution Approach 2:
The evaluation vector similarity model is designed to work with feature vectors from any speaker in a universal manner. The model learns to compare feature vectors and determine speaker identity without being speaker-specific. This universal design allows the system to accommodate new users by simply adding their feature vectors to the database, maintaining the same verification process.
3Volume of stationary object
If a compact speaker verification model is used, then the model size is reduced for client devices, but maintaining accuracy becomes more challenging
Solution Approach 1:
The system extracts only the essential components needed for speaker verification: the speaker verification model extracts features from speech, and the evaluation vector similarity model extracts speaker identity information from feature vectors. By taking out and keeping only these critical extraction and comparison functions, the system achieves compactness without sacrificing verification accuracy.
Solution Approach 2:
The system transforms raw speech data into feature vectors through the speaker verification model, changing the parameter representation from high-dimensional audio signals to condensed feature representations. This parameter transformation reduces the data size significantly while preserving the discriminative information needed for accurate speaker verification.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for speaker verification. The methods, systems, and apparatus include actions of inputting speech data that corresponds to a particular utterance to a first neural network and determining an evaluation vector based on output at a hidden layer of the first neural network. Additional actions include obtaining a reference vector that corresponds to a past utterance of a particular speaker. Further actions include inputting the evaluation vector and the reference vector to a second neural network that is trained on a set of labeled pairs of feature vectors to identify whether speakers associated with the labeled pairs of feature vectors are the same speaker. More actions include determining, based on an output of the second neural network, whether the particular utterance was likely spoken by the particular speaker.