Acoustic Model Training Data Shuffling for Information Leakage Prevention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In speech recognition, the training of acoustic models poses a risk of information leakage due to the inclusion of confidential and personal information in feature and phoneme sequences, particularly in cloud-based services where data is uploaded in its original time series format.
Innovation Solution
A shuffling process rearranges the learning data in a different order, and noise is added to the feature amounts to prevent data restoration, ensuring that even if the information leaks, it cannot be converted back to its original form, thereby enhancing the safety of model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learning data in original time series format is used for training acoustic models, then training accuracy is improved, but information leakage risk increases
Solution Approach 1:
The patent applies preliminary action by shuffling the learning data before training the acoustic model. The shuffling process rearranges the time series data in advance to eliminate temporal information that could be used to restore confidential content, while still allowing the model to learn effective acoustic patterns for accurate speech recognition.
2Measurement precision
If feature amount sequence and phoneme sequence are arranged in original time series, then speech recognition accuracy is improved, but confidential information can be restored
Solution Approach 1:
The patent applies preliminary action by shuffling the learning data before training the acoustic model. The shuffling process rearranges the time series data in advance to eliminate temporal information that could be used to restore confidential content, while still allowing the model to learn effective acoustic patterns for accurate speech recognition.
Data Source
AI summary
A learning method includes the following processes. A shuffling process acquires learning data arranged in a time series and rearranges the learning data in an order different from the order of the time series. A learning process trains an acoustic model using the learning data rearranged through the shuffling process.