Watermark-Based Synthetic Speech Detection System
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice authentication systems are vulnerable to synthetic speech and replay attacks, which can compromise the integrity of voice-based identification and authentication processes.
Innovation Solution
A centralized analysis service is implemented to detect and identify synthetic speech by embedding watermarks in synthetic speech generated by Text-to-Speech (TTS) services. These watermarks contain metadata indicating the origin, generation time, and authorization status of the synthetic speech, allowing the analysis service to differentiate between authentic and synthetic voices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If voice authentication systems use traditional voice recognition methods, then the system is simple and easy to implement, but the system becomes vulnerable to synthetic speech and replay attacks
Solution Approach 1:
The system embeds watermarks into synthetic speech during the generation process, before the speech is used for authentication. This preliminary action ensures that the watermark is already present in the audio signal when it reaches the authentication system, allowing detection without requiring additional complex processing steps at the authentication endpoint.
Solution Approach 2:
The watermark acts as an intermediary signal that carries information about the synthetic speech origin. Instead of directly analyzing the complex synthetic speech signal for authenticity, the system uses the watermark as a mediator that provides clear identification of synthetic vs. real speech, simplifying the authentication process.
2Measurement precision
If the system embeds watermarks in synthetic speech to enable detection, then the detection accuracy improves, but the complexity of the TTS service increases
Solution Approach 1:
The watermark embedding process segments the TTS generation pipeline into distinct stages: text input, watermark embedding, and speech synthesis. This segmentation allows the watermarking function to be implemented as a separate, modular component that can be independently optimized and integrated without redesigning the entire TTS service architecture.
Solution Approach 2:
The watermarking mechanism serves multiple functions simultaneously: it identifies synthetic speech origin, provides detection accuracy, and can potentially carry metadata about the generation process. This multi-functionality justifies the added complexity by delivering multiple benefits from a single integration point.
3Adaptability or versatility
If the analysis service applies multiple keys from different TTS services to detect watermarks, then the detection coverage increases, but the processing time increases
Solution Approach 1:
The system applies multiple keys to cover various TTS services, but only to the extent necessary for detection. Once a watermark is detected with a particular key, the system can stop searching other keys, avoiding unnecessary processing. This partial action approach balances detection coverage with processing efficiency.
Solution Approach 2:
The watermark detection process continues sequentially through different keys until a match is found or all keys are exhausted. This continuous action ensures that no potential synthetic speech sources are missed while maintaining a straightforward processing flow that doesn't require complex parallelization or optimization.
Data Source
AI summary
Disclosed are systems and methods including software processes executed by a server for obtaining, by a computer, an audio signal including synthetic speech, extracting, by the computer, metadata from a watermark of the audio signal by applying a set of keys associated with a plurality of text-to-speech (TTS) services to the audio signal, the metadata indicating an origin of the synthetic speech in the audio signal, and generating, by the computer, based on the extracted metadata, a notification indicating that the audio signal includes the synthetic speech.