Google LLC (20240221772). Phrase Extraction for ASR Models simplified abstract
Contents
Phrase Extraction for ASR Models
Organization Name
Inventor(s)
Ehsan Amid of Mountain View CA (US)
Om Dipakbhai Thakkar of Sunnyvale CA (US)
Rajiv Mathews of Sunnyvale CA (US)
Francoise Beaufays of Mountain View CA (US)
Phrase Extraction for ASR Models - A simplified explanation of the abstract
This abstract first appeared for US patent application 20240221772 titled 'Phrase Extraction for ASR Models
Simplified Explanation: The patent application describes a method for extracting phrases from audio data using an ASR model by modifying the audio data to obfuscate a specific phrase, then comparing the predicted transcription with the ground-truth transcription to detect leaks in the training data.
Key Features and Innovation:
- Phrase extraction method for ASR models
- Modification of audio data to obfuscate specific phrases
- Comparison of predicted transcription with ground-truth transcription
- Detection of leaks in training data
Potential Applications: This technology can be used in speech recognition systems to improve data privacy and security by detecting leaks in training data.
Problems Solved: This technology addresses the issue of unintentional leakage of sensitive information in ASR models trained on audio data.
Benefits:
- Enhanced data privacy and security in ASR models
- Improved accuracy in detecting leaks in training data
- Better control over sensitive information in audio data
Commercial Applications: Potential commercial applications include data security companies, speech recognition software developers, and organizations handling sensitive audio data.
Prior Art: Prior art related to this technology may include research on data privacy in machine learning models and techniques for detecting leaks in training data.
Frequently Updated Research: Stay updated on advancements in data privacy in machine learning models and techniques for improving the security of ASR systems.
Questions about Phrase Extraction for ASR Models: 1. How does the method obfuscate specific phrases in audio data? 2. What are the implications of detecting leaks in training data for ASR models?
Original Abstract Submitted
a method of phrase extraction for asr models includes obtaining audio data characterizing an utterance and a corresponding ground-truth transcription of the utterance and modifying the audio data to obfuscate a particular phrase recited in the utterance. the method also includes processing, using a trained asr model, the modified audio data to generate a predicted transcription of the utterance, and determining whether the predicted transcription includes the particular phrase by comparing the predicted transcription of the utterance to the ground-truth transcription of the utterance. when the predicted transcription includes the particular phrase, the method includes generating an output indicating that the trained asr model leaked the particular phrase from a training data set used to train the asr model.
