Speech Screening Models Using Transfer Learning and Segment Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing acoustic and neural network models for detecting behavioral and mental health conditions require large amounts of labeled data and have limited accuracy in predicting these conditions, with i-vector and CNN models achieving AUCs and sensitivities of only 0.60-0.64.
Innovation Solution
The proposed acoustic models utilize a transfer learning approach where an encoder is pretrained on unrelated data and a classifier is trained on labeled data, while the decoder is discarded, and segment fusion techniques are employed to improve prediction accuracy, using a combination of LSTM, RCNN, and MLP methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing acoustic models (i-vector and CNN) are used for detecting behavioral and mental health conditions, then the detection can be performed, but the accuracy is limited with AUCs and sensitivities of only 0.60-0.64
Solution Approach 1:
The encoder is pretrained on large amounts of unlabeled speech data before being used for the specific detection task. This preliminary training allows the model to learn general speech representations that improve its performance on the behavioral health detection task, resolving the contradiction by enhancing both accuracy and reliability through preparatory learning
Solution Approach 2:
The patent changes the architectural parameters and training parameters of the acoustic model by introducing a new encoder-decoder structure with specific hyperparameters (embedding dimension, hidden layer sizes, learning rates). These parameter changes enable the model to achieve higher AUC (0.75-0.79) and sensitivity, simultaneously improving measurement precision and detection reliability
2Reliability
If traditional acoustic models are trained on labeled data, then the model can be trained, but large amounts of labeled data are required which increases training time and resource requirements
Solution Approach 1:
The encoder is pretrained on large amounts of unlabeled speech data before being used for the specific detection task. This preliminary training allows the model to learn general speech representations that improve its performance on the behavioral health detection task, resolving the contradiction by enhancing both accuracy and reliability through preparatory learning
Solution Approach 2:
The model performs self-training by using its own predictions on unlabeled data to refine its representations. The encoder processes unlabeled speech data and generates representations that are then used to train the classifier, allowing the system to improve itself without requiring additional labeled data, thus reducing training time while maintaining effectiveness
Data Source
AI summary
The present disclosure provides acoustic and natural language processing (NLP) models for predicting whether a subject has a behavioral or mental health state of interest based at least in part on input speech from said subject.