ASR Memorization Scoring for Sensitive Phrase Leakage Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) models unintentionally memorize sensitive information during training, posing a risk of data leakage.
Innovation Solution
Implement a system to generate and evaluate candidate transcripts using ASR models, determining unintentional memorization by analyzing probabilities and losses, and comparing with other ASR models to identify and mitigate unintentional memorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ASR models are trained on large datasets to improve recognition accuracy, then the model's speech recognition performance is improved, but the model unintentionally memorizes sensitive information in the training data
Solution Approach 1:
The system performs preliminary actions by generating candidate transcripts and computing memorization scores before deploying the ASR model. This allows potential sensitive information leakage to be detected and addressed in advance through retraining or data removal, preventing the harmful effect from manifesting in production
Solution Approach 2:
The system implements feedback by computing memorization scores for candidate transcripts and using this information to guide further training or model updates. When high memorization scores are detected, the system can adjust training data or model parameters to reduce sensitivity to specific training examples while maintaining overall accuracy
2Measurement precision
If the ASR model is designed to be highly sensitive to specific utterances to improve accuracy on rare phrases, then recognition accuracy for rare phrases is improved, but the model increases likelihood of memorizing unique sequences
Solution Approach 1:
The system applies local quality by computing memorization scores specifically for candidate transcripts that may contain sensitive information, rather than uniformly treating all transcripts. This allows the system to maintain high sensitivity for rare phrase recognition while applying additional scrutiny and control measures only where memorization risk is detected
3Object-generated harmful factors
If the system implements comprehensive memorization detection and mitigation measures, then data security is improved, but the system complexity increases
Solution Approach 1:
The system extracts and addresses the memorization problem separately from the main ASR processing pipeline. By isolating the memorization detection and mitigation steps as distinct modules that operate on candidate transcripts, the system can implement comprehensive security measures without significantly complicating the core speech recognition functionality
Data Source
AI summary
An unintentional memorization measure can be used to determine whether an automatic speech recognition (ASR) model has unintentionally memorized one or more phrases during training of the ASR model. Various implementations include generating one or more candidate transcripts based on the vocabulary of the ASR model. For example, the system can generate a candidate transcript by appending a token of the vocabulary to a previous candidate transcript. Various implementations include processing the candidate transcript using a speech synthesis model to generate synthesized speech audio data that includes synthesized speech of the candidate transcript. Additionally or alternatively, the synthesized speech audio data can be processed using the ASR model to generate ASR output. Various implementations can include generating a loss based on comparing the ASR output and the candidate transcript.