Speech Recognition Model Weight Adjustment via Historical Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition devices using AI models often fail to accurately recognize speech inputs due to updates in the speech recognition model, leading to inconsistencies in recognizing user voice, pronunciation, intonation, accent, and dialect, resulting in inconvenient user experiences.
Innovation Solution
A method that involves training the device using a database of historical speech recognition data to adjust the weights of text candidates, allowing the device to recognize and interpret speech inputs by comparing the input with stored speech signals and updating the ASR and NLU models to maintain accurate recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the speech recognition model is updated to improve recognition accuracy, then the recognition performance is improved, but the consistency of recognizing user voice patterns deteriorates
Solution Approach 1:
The system performs preliminary actions by collecting and storing speech recognition history data before model updates occur. This historical data serves as a baseline that can be used to detect and correct inconsistencies after updates, allowing the system to proactively prepare for potential recognition drift without waiting for problems to manifest
Solution Approach 2:
The system implements feedback mechanisms by continuously comparing current speech recognition results against historical recognition patterns. When deviations are detected after model updates, the system uses this feedback to identify inconsistent recognitions and trigger retraining processes that restore recognition consistency while preserving improved accuracy
2Adaptability or versatility
If the ASR model is trained with general data to improve versatility, then the adaptability to different utterances is improved, but the accuracy in recognizing individual user patterns deteriorates
Solution Approach 1:
The system merges two distinct datasets: general speech recognition data that provides versatility across different utterances and user-specific historical data that maintains individual recognition accuracy. By combining these datasets in the retraining process, the model preserves both adaptability to various speech patterns and precision in recognizing individual user characteristics
Solution Approach 2:
The system applies local quality by differentiating between general recognition capabilities and user-specific recognition patterns. General data provides broad adaptability while user-specific historical data provides localized precision for individual users. The retraining process selectively applies appropriate data types based on the specific recognition context
Data Source
AI summary
A device for recognizing a speech input and an operating method thereof are provided. The device may be configured to: obtain one or more text candidates comprising a character string in which it is predicted that the speech input is to be converted by recognizing a speech input using an automatic speech recognition (ASR) model; extract text history information corresponding to the speech input from a database by comparing the speech input with a plurality of speech signals previously stored in the database; and perform training to adjust a weight of each of the one or more text candidates using the extracted text history information. Also, a method in which the device recognizes a speech input using an AI model may be performed.