Sequence Labeling Apparatus Using Attentional Mechanism
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sequence-labeling techniques require separate models and systems for associating speech with text and labeling, which are costly and require precise tuning, and also necessitate dividing speech into units like words or characters for effective labeling.
Innovation Solution
A sequence-labeling apparatus that uses an attentional mechanism to convert acoustic feature sequences into distributed representations, allowing for label estimation directly from these representations without requiring speech to be divided into units, by utilizing a speech distributed representation sequence converting unit, a symbol distributed representation converting unit, and a label estimation unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech is divided into units such as words or characters using a speech recognition system, then labeling can be performed on the divided units, but the system requires separate construction and optimization of speech-text association models and labeling models, increasing cost and complexity
Solution Approach 1:
The patent merges the speech-text association function and the labeling function into a single integrated model. The neural network simultaneously performs both speech-to-text association and sequence labeling, eliminating the need for separate models and their respective optimization processes, thereby reducing system complexity while maintaining labeling accuracy
Solution Approach 2:
The neural network is designed to perform multiple functions: it associates speech with text and simultaneously performs sequence labeling on the text. This multi-functional approach allows a single model to replace what would traditionally require separate specialized models, reducing overall system complexity
2Reliability
If separate models for speech-text association and labeling are constructed, then each model can be optimized for its specific function, but the cost of construction and tuning increases significantly
Solution Approach 1:
By combining speech-text association and labeling into one neural network model, the patent eliminates the need to separately construct and tune multiple models. The single model is trained end-to-end on both tasks simultaneously, reducing construction and tuning costs while maintaining reliable association accuracy through joint optimization
3Ease of operation
If a speech recognition system is constructed to divide acoustic signals into units, then labeling can be performed on the divided units, but the process requires advance construction and optimization of the speech recognition system
Solution Approach 1:
The patent performs speech-text association and labeling in a single integrated processing step rather than requiring preliminary speech recognition to divide speech into units first. The neural network directly processes the acoustic signal and simultaneously performs association and labeling, eliminating the time-consuming preliminary division step while maintaining ease of operation
Solution Approach 2:
The integrated neural network performs speech-text association and labeling in continuous operation without interrupting the processing flow for separate speech recognition and unit division. This continuous processing reduces overall processing time while keeping the labeling process simple and efficient
Data Source
AI summary
Without dividing speech into a unit such as a word or a character, text corresponding to the speech is labeled. A speech distributed representation sequence converting unit 11 converts an acoustic feature sequence into a speech distributed representation. A symbol distributed representation converting unit 12 converts each symbol included in the symbol sequence corresponding to the acoustic feature sequence into a symbol distributed representation. A label estimation unit 13 estimates a label corresponding to the symbol from the fixed-length vector of the symbol generated using the speech distributed representation, the symbol distributed representation, and fixed-length vectors of previous and next symbols.