Voice Recognition Model Segmentation for O2O Services
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice recognition models for online to offline services are not trained in real-time, leading to inaccurate voice recognition results when users request services via voice, as they are typically trained offline and updated periodically.
Innovation Solution
A system that includes a processor and storage medium for receiving voice requests from user terminals, obtaining customized recognition models trained on user-specific data and general recognition models trained from public data, and determining literal destinations by comparing results from both models to improve accuracy and update models in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If voice recognition models are trained offline and updated periodically, then the system complexity is reduced and ease of operation is improved, but the measurement precision of voice recognition accuracy deteriorates
Solution Approach 1:
The voice recognition model is segmented into two distinct components: a general recognition model trained offline on public data and a customized recognition model trained online on user-specific data. This segmentation allows the system to maintain high accuracy through real-time customization while keeping the overall architecture manageable by separating training responsibilities.
Solution Approach 2:
The general recognition model is pre-trained offline using public data before the user actually needs voice recognition services. This preliminary action provides a solid baseline that can be quickly refined with user-specific data without requiring the entire model to be retrained from scratch, thus improving accuracy without proportionally increasing system complexity.
2Measurement precision
If voice recognition models are trained in real-time using user-specific data, then the measurement precision of voice recognition accuracy is improved, but the loss of time for model training increases
Solution Approach 1:
By dividing the model into general and customized components, the system only needs to train the customized recognition model using user-specific historical data rather than retraining the entire model. This segmentation significantly reduces the training time required while still achieving high accuracy for user-specific voice recognition needs.
Solution Approach 2:
Instead of training the complete voice recognition model from scratch for each user (excessive action), the system trains only the customized portion of the model using user-specific historical data (partial action). This partial training approach achieves sufficient accuracy for the user's specific needs without the time cost of full model retraining.
3Adaptability or versatility
If a customized recognition model is trained for each user, then the adaptability of the system to user-specific pronunciation patterns is improved, but the device complexity increases
Solution Approach 1:
The model is segmented into a universal general recognition model and user-specific customized recognition models. This segmentation enables the system to adapt to individual user pronunciation patterns through lightweight customized models while avoiding the complexity of managing entirely separate models for each user, as the general model serves as a shared foundation.
Solution Approach 2:
The system merges the general recognition model with customized recognition models to form a complete voice recognition system. This combining approach allows the system to leverage both universal knowledge from the general model and user-specific patterns from the customized model, achieving high adaptability without the complexity of completely independent models for each user.
Data Source
AI summary
Systems and methods for providing an online to offline service in response to a voice request from a user terminal are provided. A method includes: receiving a voice request from a user terminal; in response to the voice request, updating a customized recognition model trained using data of a plurality of points of interest associated with the user terminal; obtaining a general recognition model trained using data from general public; determining a literal destination associated with the voice request based at least on the voice request, the customized recognition model and the general recognition model.