Dynamic Time Warping for Voice Spoofing Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice utterance comparison methods fail to reliably identify speakers and detect spoofing, such as cut and paste spoofing, due to variations in pronunciation and speech rate, and the inability to distinguish between genuine and fake voice utterances.
Innovation Solution
The method involves extracting features from voice utterances, applying dynamic time warping to align time-dependent characteristics, and calculating a total distance measure using multiple features and distance functions like Euclidean, Mahalanobis, and Cosine distances to differentiate between genuine and spoofed voices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If dynamic time warping is applied to align time-dependent characteristics, then speaker identification efficiency is improved, but the ability to detect cut and paste spoofing deteriorates
Solution Approach 1:
The patent segments the voice utterance into multiple features (spectral features, temporal features, prosodic features) and applies dynamic time warping to each feature independently. This segmentation allows the system to maintain flexibility for speaker identification while detecting inconsistencies in individual features that indicate spoofing.
Solution Approach 2:
The patent applies different evaluation criteria to different features. Some features are evaluated for their overall alignment quality (supporting speaker identification), while other features are specifically evaluated for abrupt temporal changes or inconsistencies (supporting spoofing detection). This local differentiation resolves the contradiction by optimizing for both goals simultaneously.
2Measurement precision
If multiple features are extracted and compared, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent divides the voice comparison task into multiple independent feature extraction and comparison streams. Each feature type (spectral, temporal, prosodic) is processed separately through its own distance function, allowing parallel computation that manages complexity while maintaining high measurement precision through multi-feature integration.
Solution Approach 2:
The patent employs a universal dynamic time warping framework that handles multiple feature types through a common alignment mechanism. This multi-functional approach allows the same core algorithm to process different feature types, reducing overall system complexity despite the increased number of features being compared.
3Measurement precision
If distance measures are minimized to align voice utterances, then speaker identification accuracy is improved, but the ability to detect temporal inconsistencies in spoofed voices deteriorates
Solution Approach 1:
The patent applies different evaluation criteria to different features. Some features are evaluated for their overall alignment quality (supporting speaker identification), while other features are specifically evaluated for abrupt temporal changes or inconsistencies (supporting spoofing detection). This local differentiation resolves the contradiction by optimizing for both goals simultaneously.
Solution Approach 2:
The system uses the results from multiple distance functions as feedback to evaluate both alignment quality and temporal consistency. The combined evaluation provides feedback that simultaneously optimizes speaker identification accuracy and detects temporal inconsistencies, allowing the system to achieve both objectives without compromising either.
Data Source
AI summary
The invention refers to a method for comparing voice utterances, the method comprising the steps: extracting a plurality of features (201) from a first voice utterance of a given text sample and extracting a plurality of features (201) from a second voice utterance of said given text sample, wherein each feature is extracted as a function of time, and wherein each feature of the second voice utterance corresponds to a feature of the first voice utterance; applying dynamic time warping (202) to one or more time dependent characteristics of the first and/or second voice utterance e.g. by minimizing one or more distance measures, wherein a distance measure is a measure for the difference of a time dependent characteristic of the first voice utterance and a corresponding time dependent characteristic of the second voice utterance, and wherein a time dependent characteristic of a voice utterance is a time dependent characteristic of either a single feature or a combination of two or more features; calculating a total distance measure (203), wherein the total distance measure is a measure for the difference between the first voice utterance of the given text sample and the second voice utterance of said given text sample, wherein the total distance measure is calculated based on one or more pairs of said time dependent characteristic, and wherein a pair of time dependent characteristic is calculate total composed of a time dependent characteristic of the first or second voice utterance and of a dynamically time warped (202) time dependent characteristic of the respectively second or first voice utterance, or wherein a pair of time dependent characteristic is composed of a dynamically time warped (202) time dependent characteristic of the first voice utterance and of a dynamically time warped (202) time dependent characteristic of the second voice utterance.