Speaker-Independent Audio Embeddings for Spoof-Resistant Verification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speaker verification systems are susceptible to voice spoofing and background noise, and rely on unreliable telephony identifiers, necessitating a means to verify audio sources based on speaker-independent characteristics.
Innovation Solution
Systems and methods utilizing machine learning models, including Gaussian Mixture Models and neural networks, to extract speaker-independent embeddings and deep-phoneprints (DP vectors) from audio signals, which can differentiate between various audio aspects such as device type, microphone type, geographical location, and network type, and detect spoofing services.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speaker-dependent voice biometrics are used for authentication, then speaker identification accuracy is improved, but the system becomes vulnerable to voice spoofing and deepfakes
Solution Approach 1:
The patent segments the audio signal analysis into two independent components: speaker-dependent voice biometrics (for identifying the speaker) and speaker-independent audio characteristics (for verifying audio source authenticity). By analyzing device type, microphone type, network type, and audio quality separately from the voice content, the system can detect spoofing attempts while maintaining accurate speaker identification.
Solution Approach 2:
The patent introduces speaker-independent audio characteristics as an intermediary verification layer between the voice biometric authentication and the final authentication decision. This intermediary component analyzes audio properties that are independent of the speaker's voice, such as device characteristics and audio quality metrics, providing an additional security layer that is not vulnerable to voice spoofing.
2Ease of operation
If telephony identifiers like caller ID are used for verification, then authentication simplicity is improved, but reliability deteriorates due to spoofing capabilities
Solution Approach 1:
The patent replaces the traditional telephony identifier verification mechanism with an audio signal analysis mechanism. Instead of relying on Caller ID, ANI, or other telephony metadata that can be spoofed, the system uses machine learning models to analyze actual audio characteristics captured during the call, substituting a vulnerable mechanical/protocol-based system with an acoustic analysis-based system.
3Reliability
If speaker-independent audio characteristics are analyzed, then vulnerability to voice spoofing is reduced, but system complexity increases
Solution Approach 1:
The patent segments the complex audio analysis task into multiple specialized machine learning models, each dedicated to analyzing specific audio characteristics (device type, microphone type, network type, audio quality). This modular approach manages complexity by dividing the overall system into independent, manageable components that can be trained and deployed separately.
Solution Approach 2:
The patent creates a universal authentication framework where the same speaker-independent audio characteristic analysis can serve multiple functions: detecting voice spoofing, identifying device type, assessing audio quality, and verifying audio source authenticity. This multi-functionality justifies the system complexity by providing multiple security and analytical capabilities through a single integrated approach.
Data Source
AI summary
Embodiments described herein provide for audio processing operations that evaluate characteristics of audio signals that are independent of the speaker's voice. A neural network architecture trains and applies discriminatory neural networks tasked with modeling and classifying speaker-independent characteristics. The task-specific models generate or extract feature vectors from input audio data based on the trained embedding extraction models. The embeddings from the task-specific models are concatenated to form a deep-phoneprint vector for the input audio signal. The DP vector is a low dimensional representation of the each of the speaker-independent characteristics of the audio signal and applied in various downstream operations.