Personalized Speech Recognition Model Using Scaling Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition models lack personalization, leading to suboptimal performance for individual users due to their reliance on general speech data, which fails to account for unique voice characteristics and regional dialects.
Innovation Solution
A processor-implemented method that generates a personalized speech recognition model by modifying a base model using user-specific utterance data and statistical information, such as mean and variance of scaling vectors, to adjust the output of hidden layers and reflectance rates in the loss function, thereby adapting to individual speech patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general speech recognition model is used, then the model can recognize speech from multiple speakers, but the recognition accuracy for individual users is suboptimal
Solution Approach 1:
The speech recognition model is segmented into a base model and user-specific scaling vectors. The base model handles general speech recognition for multiple speakers, while separate scaling vectors are introduced for each user to capture individual voice characteristics. This segmentation allows the system to maintain versatility across users while achieving personalized accuracy through user-specific adaptations.
2Measurement precision
If user-specific training data is collected, then personalization accuracy improves, but the quantity of required data increases
Solution Approach 1:
Instead of training entire models with large datasets, the invention changes the approach by training only scaling vectors with minimal user data. The scaling vectors are initialized from pre-trained base models and then fine-tuned using small amounts of user utterance data. This parameter change from full model training to scaling vector optimization enables effective personalization with limited data resources.
3Measurement precision
If scaling vectors are trained from scratch, then personalized performance improves, but training time and computational resources increase
Solution Approach 1:
The base model is pre-trained on large datasets before personalization. Scaling vectors are initialized with weights from this pre-trained model rather than starting from random values. This preliminary action of pre-training the base model and initializing scaling vectors accordingly significantly reduces the training time required for personalization, as the scaling vectors only need to be fine-tuned with user-specific data rather than trained from scratch.
Data Source
AI summary
A processor-implemented method of personalizing a speech recognition model includes: obtaining statistical information of first scaling vectors combined with a base model for speech recognition; obtaining utterance data of a user; and generating a personalized speech recognition model by modifying a second scaling vector combined with the base model based on the utterance data of the user and the statistical information.