Dual-Encoder Speech Recognition for Context and Foreign Words
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice recognition technologies struggle with accurately recognizing foreign words and fail to fully reflect the context of voice signals, leading to mismatched recognition results due to context-dependent encoders being biased by previous words.
Innovation Solution
The electronic device classifies encoders into context-independent and context-dependent types, using a common encoder to learn vectors suitable for both, followed by individual encoders and decoders to process phoneme and sub-word sequences, with spell and named entity corrections based on phoneme sequences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a context-dependent encoder is used to fully reflect the context of the voice signal, then the context utilization is improved, but the recognition accuracy of foreign words deteriorates due to strong bias towards previous words
Solution Approach 1:
The encoder is segmented into two independent components: a context-independent encoder that processes each voice signal section independently to maintain accuracy for foreign words, and a context-dependent encoder that captures contextual information. These separate encoders work in parallel to address the contradiction between context utilization and recognition accuracy.
Solution Approach 2:
Different encoding strategies are applied to different aspects of the voice signal processing. The context-independent encoder provides localized, section-by-section processing for accurate foreign word recognition, while the context-dependent encoder provides global contextual information. Each encoder has specialized quality characteristics suited to its specific function.
2Productivity
If a limited section of the voice signal is used for voice recognition, then the processing speed is improved, but the recognition accuracy deteriorates due to incomplete context reflection
Solution Approach 1:
The voice signal is segmented into multiple sections that are processed independently by the context-independent encoder, enabling parallel processing and improved speed. Meanwhile, the context-dependent encoder aggregates information from these sections to provide complete contextual understanding, resolving the contradiction between processing speed and recognition accuracy.
Solution Approach 2:
The solution adds a temporal dimension to the processing architecture by using the context-dependent encoder to incorporate information from before and after the current time point. This dimensional expansion allows the system to maintain both fast processing of individual sections and comprehensive contextual analysis.
Data Source
AI summary
Provided are an electronic device and a method of controlling an electronic device. The electronic device includes: a memory storing at least one instruction; and at least one processor configured to execute the at least one instruction, wherein one or more of the at least one processor is configured to: acquire a first vector corresponding to each of a plurality of sections of a voice signal by inputting the voice signal to a common encoder based on acquiring the voice signal; acquire a second vector corresponding to each of the plurality of sections and independent on a context of the voice signal by inputting the first vector into a first individual encoder; acquire a phoneme sequence corresponding to the second vector by inputting the second vector into a first decoder; acquire a third vector corresponding to at least two sections among the plurality of sections and dependent on the context of the voice signal by inputting the first vectors into a second individual encoder; acquire a sub-word sequence corresponding to the third vector by inputting the third vector into a second decoder; and acquire text information corresponding to the plurality of sections by correcting the sub-word sequence based on the phoneme sequence, through a text information acquisition module.