Speech Recognition Model Training With Attribute-Guided Expert Assignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition models based on Mixture-Of-Experts (MoE) suffer from low accuracy and high training costs due to unsupervised sample assignment, leading to inefficient use of resources and suboptimal feature extraction.
Innovation Solution
A supervised training method for speech recognition models that classifies speech samples under target attributes, adjusting network parameters based on attribute category differences and using a combination of encoding and decoding networks to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If unsupervised training with random sample assignment is used, then the training process is simpler to implement, but the speech recognition accuracy rate decreases
Solution Approach 1:
The patent introduces supervised training with attribute category labels as feedback signals. The encoding network receives speech samples and their corresponding attribute categories, producing encoded features that are then used for training the recognition model. This feedback mechanism enables the model to learn accurate mappings between speech attributes and categories, significantly improving recognition accuracy compared to unsupervised random assignment.
Solution Approach 2:
The patent performs preliminary classification of speech samples into attribute categories before training the recognition model. By pre-organizing the training data with clear attribute labels and assigning samples to appropriate experts based on their characteristics, the system creates a structured learning foundation that improves both training efficiency and final model accuracy.
2Adaptability or versatility
If a large number of samples and experts are used for training, then the model can handle more complex patterns, but the training cost becomes very high
Solution Approach 1:
The patent divides the speech recognition task into separate functional modules: an encoding network that processes speech samples and extracts features, and a recognition model that performs the actual recognition. This segmentation allows each component to be optimized independently and enables efficient training with fewer total samples by leveraging the pre-extracted encoded features.
Solution Approach 2:
The encoding network serves multiple functions: it processes speech samples, extracts relevant features, classifies attribute categories, and prepares data for the recognition model. This multi-functionality reduces the need for separate specialized components and decreases overall training requirements while maintaining high model adaptability and versatility.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Disclosed is a speech recognition method. The method comprises: acquiring a speech to be recognized and a trained speech recognition model, the speech recognition model comprising an encoding network and a decoding network (S11); performing classification under a target speech attribute to obtain a predicted attribute category to which said speech belongs, and performing encoding based on the predicted attribute category under the target speech attribute to obtain a first encoding feature (S12); and decoding the first encoding feature based on the decoding network to obtain a recognition text of said speech (S13), wherein the speech recognition model is adjusted at least based on a first loss, and the first loss represents the difference, under the target speech attribute, between a preset attribute category of a speech sample annotation and a sample attribute category obtained after the recognition of the speech recognition model.