Speech Recognition Using Dual Decoding to Reduce Information Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems suffer from information loss during model interactions, leading to limited recognition accuracy due to sequential mapping relationships between speech features, phonemes, and strings.
Innovation Solution
A speech recognition method that performs feature extraction on speech content to obtain an intermediate feature, which is then processed using both an attention mechanism and pronunciation-based feature mapping to generate two word graph networks, allowing for improved candidate sequence prediction and enhanced accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a conventional speech recognition system with multiple models is used, then the system can process speech information through sequential mapping relationships, but information loss occurs during information transfer between models, resulting in limited recognition accuracy
Solution Approach 1:
The patent merges the acoustic model and language model into a unified neural network architecture with shared layers. The intermediate features extracted by the shared acoustic feature extraction layer are simultaneously fed to both the attention mechanism decoder and the pronunciation-based decoder, eliminating information loss through sequential model processing while maintaining the functional benefits of both modeling approaches.
Solution Approach 2:
The patent segments the speech recognition system into distinct processing paths: one path uses an attention mechanism to decode intermediate features into candidate words, while another path uses pronunciation-based feature mapping. These segmented paths operate in parallel and their results are combined, allowing each path to specialize while avoiding information loss in sequential processing.
2Adaptability or versatility
If multiple models are used in the speech recognition system, then comprehensive processing can be achieved, but the complexity of the system increases due to information transfer between various models
Solution Approach 1:
The patent implements a universal acoustic feature extraction layer that serves multiple functions: it extracts features for the attention mechanism decoder, provides features for the pronunciation-based decoder, and maintains compatibility with both processing paths. This multi-functional design achieves comprehensive processing capability while reducing system complexity by eliminating redundant feature extraction components.
3Measurement precision
If sequential mapping relationships between speech features, phonemes, and strings are used, then the conventional speech recognition system can operate, but recognition accuracy is limited due to information loss
Solution Approach 1:
The patent performs preliminary action by extracting acoustic intermediate features in advance through the shared acoustic feature extraction layer before the decoding stage. These pre-extracted intermediate features are then simultaneously available to both the attention mechanism and pronunciation-based processing paths, eliminating the need for repeated feature extraction and reducing information transfer time while improving recognition accuracy.
Data Source
AI summary
A computer device acquires speech content. The device performs feature extraction on the speech content to obtain an intermediate feature. The intermediate feature is used for indicating an audio expression characteristic of the speech content. The device decodes the intermediate feature based on an attention mechanism to obtain a first word graph network. The device performs feature mapping on the intermediate feature based on pronunciation of the speech content to obtain a second word graph network. The device determines a recognition result of the speech content according to candidate word connection relationships indicated by the first word graph network and the second word graph network.