PII-Safe Speech Recognition Training Set Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models for speech recognition face challenges in maintaining accuracy over time due to changes in input data distributions, and they require large amounts of diverse data to improve performance, while also needing to handle private information that must be protected.
Innovation Solution
A system and method for continuous re-training of a speech recognition machine learning model, which involves selecting and processing audio files from call centers to create an optimized training set by filtering out private information and fragments with low confidence scores, and then manually transcribing and re-training the model using the filtered data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a speech recognition model is continuously trained with new data to improve accuracy, then the model's accuracy improves, but the risk of including private information in the training data increases
Solution Approach 1:
The system performs preliminary filtering of audio data to detect and remove segments containing private information before the data is used for model training. This proactive approach prevents private information from entering the training pipeline, resolving the contradiction between using real-world data for improvement and protecting privacy.
Solution Approach 2:
The patent introduces an intermediary detection system that acts as a mediator between the audio data source and the training process. This intermediary layer analyzes audio segments for private information using multiple criteria (speaker identification, context analysis, confidence scoring) and selectively blocks problematic segments, enabling both accuracy improvement and privacy protection.
2Adaptability or versatility
If the model is re-trained frequently to adapt to new data distributions, then the model's adaptability improves, but the computational resources and time required increase
Solution Approach 1:
Instead of re-training the model with all available data, the system selectively identifies and extracts only the most valuable audio segments that contain novel linguistic patterns or represent new data distributions. This partial action approach achieves adaptability improvement with reduced computational overhead and faster re-training cycles.
Solution Approach 2:
The system dynamically adjusts training parameters such as learning rate, batch size, and selection criteria based on the characteristics of incoming data and model performance metrics. This enables efficient adaptation to new data distributions by optimizing the re-training process itself, reducing the time and resources required.
3Reliability
If strict filtering is applied to remove all segments with private information, then privacy protection is ensured, but the quantity of training data decreases
Solution Approach 1:
The system applies different filtering strictness levels to different segments of audio data based on local characteristics. Segments with high confidence of containing private information are strictly filtered, while segments with low confidence or ambiguous content are retained after additional analysis. This localized quality control maintains privacy protection while preserving more training data.
Solution Approach 2:
The system performs preliminary analysis and scoring of audio segments to identify those most likely to contain private information before applying strict filtering. By pre-screening data and prioritizing removal of high-risk segments, the system achieves strong privacy protection with minimal loss of usable training data.
Data Source
Figure 1
Figure 2
AI summary
A method comprising receiving, as input, one or more audio files; applying a trained speech recognition algorithm to said one or more audio files, to obtain textual output corresponding to each of said one or more audio files; extracting, based on said textual output, from each of said one or more audio files, one or more portions having a specified syntactic pattern; selecting a subset of said portions based on at least one of: (i) a content of said textual output associated with each of said portions, (ii) a duration of each of said portions, and (iii) a confidence score assigned by said trained speech recognition algorithm to said obtained textual output; receiving, as input, transcriptions of each of said portions; generating a re-training set comprising: (iv) said portions in said subset, and (iv) said transcriptions; and re-training said trained speech recognition algorithm on said re-training set.