Depersonalizing Text Corpora for Speech Recognition Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face challenges in training due to the presence of personally identifiable information (PII) in in-domain data, which cannot be retained due to data retention restrictions, hindering the effectiveness of language and acoustic models.
Innovation Solution
A method and system for depersonalizing a text corpus by replacing PII with class labels, filtering out identifiable information, and shredding the data into non-overlapping segments and audio features, allowing for the training of language and acoustic models without retaining original messages, thus preventing PII reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If in-domain data containing PII is retained for training, then model training effectiveness is improved, but data retention restrictions are violated
Solution Approach 1:
The patent extracts and removes personally identifiable information from the training corpus, separating useful training data from harmful PII. This allows the system to retain beneficial in-domain data for model training while eliminating compliance violations by removing identifiable personal information before training processes.
2Object-affected harmful factors
If original messages are deleted to prevent PII reconstruction, then data privacy is protected, but training data availability is reduced
Solution Approach 1:
The patent segments the original messages into smaller textual units such as n-grams or phrases, maintaining sufficient training data for model effectiveness while preventing reconstruction of complete original messages. This segmentation preserves linguistic patterns and statistical information needed for training without enabling PII recovery.
Solution Approach 2:
The patent creates processed copies of the original training data through depersonalization and segmentation. These copies retain the statistical and linguistic properties necessary for effective model training while being fundamentally different from the original messages, thus preventing reconstruction of sensitive information.
3Object-affected harmful factors
If data is depersonalized and segmented, then PII is removed and privacy is protected, but data utility for training may be reduced
Solution Approach 1:
The patent applies partial depersonalization by removing only the personally identifiable portions of data while retaining the majority of the textual content. This partial action maintains data utility for training language models and acoustic models without requiring complete removal of all original information, thus balancing privacy protection with training effectiveness.
Data Source
AI summary
Training speech recognizers, e.g., their language or acoustic models, using actual user data is useful, but retaining personally identifiable information may be restricted in certain environments due to regulations. Accordingly, a method or system is provided for enabling training of a language model which includes producing segments of text in a text corpus and counts corresponding to the segments of text, the text corpus being in a depersonalized state. The method further includes enabling a system to train a language model using the segments of text in the depersonalized state and the counts. Because the data is depersonalized, actual data may be used, enabling speech recognizers to keep up-to-date with user trends in speech and usage, among other benefits.