Speech Recognition Device Using Dual Encoder Decoder Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies using end-to-end (E2E) models face challenges in robustness, with online recognition offering fast response but poorer performance compared to offline recognition, and vice versa, depending on the utterance.
Innovation Solution
A speech recognition device and method that integrates online and offline processing using a dual encoder and decoder architecture, where the first encoder and decoder perform online processing, the second encoder and decoder perform offline processing, and an integration unit rescoringly combines their outputs for improved robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If online recognition is used, then response speed is improved, but recognition performance deteriorates
Solution Approach 1:
The patent combines online and offline recognition models into a unified system. The integration unit merges the output from the online decoder (fast response) and offline decoder (high performance) to produce a final recognition result, thereby achieving both speed and accuracy simultaneously
2Reliability
If offline recognition is used, then recognition performance is improved, but response speed deteriorates
Solution Approach 1:
The system integrates offline recognition (high performance) with online recognition (fast response) through the integration unit. The offline decoder provides accurate recognition results while the online component maintains fast response capability, achieving a balance between performance and speed
3Ease of manufacture
If fixed block length is used in learning, then training simplicity is improved, but robustness deteriorates
Solution Approach 1:
The patent introduces dynamic block length selection during the learning phase. The learning unit randomly selects different block lengths for training the encoder and decoder, making the model adaptable to various utterance lengths and improving robustness without significantly complicating the training process
Data Source
AI summary
A speech recognition device is a speech recognition device for performing speech recognition using an end-to-end model and includes an encoder that is a model that converts an input speech signal into a characteristic expression, a decoder that is a model that converts speech data into text using an output of the encoder, and a learning unit that randomly selects a block length of the speech signal input to the encoder and causes the encoder and the decoder to learn.