Voice Discrimination Using ML Classifier for Virtual Assistant Security
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems for virtual assistants struggle to distinguish between direct human voices and machine-generated voices, leading to unintended responses to non-human audio inputs, such as advertisements or malicious commands, which can result in unauthorized actions.
Innovation Solution
A method and system utilizing a machine learning classifier to differentiate between direct and machine-generated human voices by capturing audio samples from both sources and extracting discriminative features, allowing selective response generation based on the identified origin of the voice input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If virtual assistant systems respond to all voice inputs, then user convenience is improved, but security deteriorates due to unintended responses to machine-generated voices
Solution Approach 1:
The patent introduces an intermediary classification system between the voice input and the virtual assistant response. The system uses acoustic feature analysis and machine learning classifiers as intermediaries to distinguish human voices from machine-generated voices, allowing the virtual assistant to selectively respond only to genuine human inputs while maintaining convenience for legitimate users and preventing unauthorized automated commands
2Measurement precision
If acoustic fingerprinting algorithms are used to identify voice sources, then response accuracy is improved, but applicability deteriorates when source audio samples are unavailable
Solution Approach 1:
The patent replaces the mechanical approach of acoustic fingerprinting (which requires pre-existing reference samples) with a signal processing approach that analyzes intrinsic acoustic characteristics of the input voice itself. By examining features like formant frequencies, spectral envelope, and temporal patterns directly in the incoming audio, the system achieves accurate human vs. machine voice differentiation without needing prior knowledge of the specific voice source
Solution Approach 2:
The patent transforms the problem from source identification (requiring reference samples) to characteristic analysis (using intrinsic parameters). By changing the analysis parameters from comparing against stored fingerprints to evaluating real-time acoustic features such as spectral tilt, harmonics-to-noise ratio, and formant structure, the system becomes adaptable to any voice input regardless of whether source samples are available
3Object-affected harmful factors
If the system discriminates between human and machine voices, then security is improved, but system complexity increases due to additional processing requirements
Solution Approach 1:
The patent applies partial action by implementing a tiered processing approach. Instead of performing full spectral analysis on all audio inputs, the system first applies lightweight acoustic feature extraction that quickly identifies probable machine-generated voices using simple metrics like spectral flatness and zero-crossing rate. Only inputs that pass this initial screening undergo more sophisticated analysis, reducing overall computational complexity while maintaining security
Solution Approach 2:
The system employs self-service through adaptive thresholding and confidence-based decision making. The classification algorithm automatically adjusts its decision thresholds based on the distribution of features in the input population, and when classification confidence is high, the system makes immediate decisions without requiring additional verification steps. This self-adjusting behavior reduces processing complexity while maintaining high security standards
Data Source
AI summary
Discriminating between direct and machine-generated human voices is disclosed. A directly-generated voice audio sample from a human utterance and a machine-generated voice audio sample outputted by a loudspeaker from a pre-recording of another human utterance are captured on a microphone. Discriminative features between the directly-generated voice audio sample and the machine-generated voice audio sample are extracted with a machine learning classifier. A response to a command in the captured directly-generated voice audio sample or the captured machine-generated voice audio sample may be selectively generated.