Speech Synthesizer Class Probability Correction for Reading Break Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The performance of synthesized speech models deteriorates due to class imbalance in training data, leading to unnatural reading breaks and user discomfort when listening to synthesized speech.
Innovation Solution
A speech synthesizer adjusts the class classification probabilities based on the ratios of words classified into minor and major classes, using a processor to change the first class classification probability set to a second class classification probability set, which includes a minor class with the smallest count, to improve reading break prediction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training data with class imbalance is used for synthesized speech model learning, then the model can be trained with available data, but the reading break prediction performance deteriorates
Solution Approach 1:
The patent modifies the class classification probabilities by applying correction factors based on the ratio between minor and major class counts. This changes the parameter distribution in the training data to balance the representation of different reading break types, thereby improving prediction performance while using the same training dataset.
Solution Approach 2:
The patent applies a correction factor that may exceed 1.0 for minor classes, effectively oversampling or overweighting these underrepresented classes during model training. This partial action on specific classes (rather than all classes) addresses the imbalance without requiring complete data redistribution.
2Ease of manufacture
If the synthesized speech model is trained with uncorrected class probabilities, then the training process is simpler, but the output speech becomes unnatural and users feel uncomfortable
Solution Approach 1:
The patent performs preliminary correction of class classification probabilities before model training by calculating correction factors based on class distribution ratios. This preliminary action ensures that the training data is pre-balanced, leading to more natural synthesized speech output without complicating the overall training workflow.
3Measurement precision
If class classification probabilities are corrected based on minor class ratios, then reading break prediction performance improves, but the processing complexity increases
Solution Approach 1:
The system automatically calculates correction factors using the observed class distribution in the training data itself. The minor class count and major class count are derived from the same dataset, allowing the system to self-correct the imbalance without external intervention or complex manual adjustment processes.
Data Source
AI summary
A speech synthesizer using artificial intelligence includes a memory configured to store a first ratio of a word classified into a minor class among a plurality of classes, a second ratio of the word which is not classified into the minor class, and a synthesized speech model and a processor configured to change a first class classification probability set of the word to a second class classification probability set, based on the first ratio, the second ratio and the first class classification probability set, and learn the synthesized speech model using the changed second class classification probability set.