17643071. ADVERSARIAL SPEECH-TEXT PROTECTION AGAINST AUTOMATED ANALYSIS simplified abstract (INTERNATIONAL BUSINESS MACHINES CORPORATION)
ADVERSARIAL SPEECH-TEXT PROTECTION AGAINST AUTOMATED ANALYSIS
Organization Name
INTERNATIONAL BUSINESS MACHINES CORPORATION
Inventor(s)
ADVERSARIAL SPEECH-TEXT PROTECTION AGAINST AUTOMATED ANALYSIS - A simplified explanation of the abstract
This abstract first appeared for US patent application 17643071 titled 'ADVERSARIAL SPEECH-TEXT PROTECTION AGAINST AUTOMATED ANALYSIS
Simplified Explanation
The patent application describes a method, system, and computer program for processing audio signals containing speech data and generating text data from it. It also involves identifying vulnerable portions of the text data and applying adversarial text to make it more robust. Adversarial noise is then generated and applied to the speech data.
- The method involves processing audio signals containing speech data and converting it into text data.
- Vulnerable portions of the text data are identified.
- Adversarial text is applied to the vulnerable portions of the text data to make it more robust.
- Adversarial noise is generated based on the robust text data.
- The generated adversarial noise is applied to the speech data.
Potential Applications
This technology has potential applications in various fields, including:
- Speech recognition systems
- Transcription services
- Voice assistants
- Call center operations
- Language translation services
Problems Solved
The technology addresses the following problems:
- Vulnerabilities in text data generated from speech signals
- Adversarial attacks on speech recognition systems
- Improving the robustness and accuracy of speech-to-text conversion
Benefits
The technology offers several benefits:
- Enhanced security and protection against adversarial attacks
- Improved accuracy and reliability of speech-to-text conversion
- Increased robustness of text data generated from speech signals
- Potential for better performance in speech recognition systems.
Original Abstract Submitted
A method, system, and computer program product are disclosed. The method includes processing an audio signal that includes speech data and transcribing the speech data to generate text data. The method also includes identifying a vulnerable portion of the text data and, in response, applying adversarial text to the text data to generate robust text data. Adversarial noise corresponding to the robust text data is generated and applied to the speech data.
