Noisy Speech Recognition via Dynamic Lexicon Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional automatic speech recognition systems produce noisy outputs with high error rates, and existing grammars are domain-specific, difficult to modify, and not suitable for noisy, uncertain data, limiting their effectiveness in interpreting and acting upon noisy textual inputs.
Innovation Solution
A method combining linguistics and computer science techniques using thematic relation patterns, homonym tables, dynamic lexicons, and regular expressions to generate and score all possible interpretations of noisy input sequences, allowing for interactive modification and parallel processing to select the best interpretation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional grammars are used for speech recognition, then domain-specific accuracy is improved, but adaptability to noisy and uncertain data deteriorates
Solution Approach 1:
The patent implements dynamic lexicons that can be updated and modified during runtime based on observed data patterns and user feedback. Unlike static traditional grammars, the system adapts its vocabulary and thematic patterns dynamically to handle noisy input, allowing the recognition model to evolve and improve its noise tolerance while maintaining domain-specific accuracy.
Solution Approach 2:
The system changes parameters by using multiple thematic patterns with varying specificity levels and combining them with probabilistic methods. It adjusts the weight and application of different grammatical rules based on the noise level detected in input data, allowing flexible parameter adjustment to balance between strict domain compliance and robustness to noise.
2Measurement precision
If traditional grammars are used, then domain-specific recognition is improved, but ease of modification deteriorates
Solution Approach 1:
The patent segments the grammar system into modular thematic patterns that can be independently defined, stored in tables, and modified without affecting the entire grammar structure. This segmentation allows users to easily add, remove, or update specific thematic patterns and lexicon entries while maintaining the overall domain-specific recognition framework.
Solution Approach 2:
The system uses a universal thematic pattern framework that can handle multiple domains and languages through a common structure. The same basic pattern matching mechanism works across different domains by simply changing the thematic pattern tables and lexicons, making the system both domain-specific in capability and universally adaptable in implementation.
3Adaptability or versatility
If all possible interpretations are generated for noisy input, then noise tolerance is improved, but processing time increases
Solution Approach 1:
The system performs preliminary action by pre-compiling thematic patterns, lexicons, and probability models before processing actual speech input. This pre-processing creates a ready-to-use framework that enables rapid matching during runtime, allowing the system to generate and evaluate multiple interpretations quickly without incurring excessive processing delays during actual use.
Solution Approach 2:
The patent replaces traditional sequential mechanical processing with parallel processing and probabilistic methods. Multiple interpretations are generated and evaluated simultaneously using parallel computation, and probabilistic scoring replaces step-by-step deterministic validation, significantly reducing the time required to process noisy input while maintaining high noise tolerance.
Data Source
AI summary
This invention maps possibly noisy digital input from any of a number of different hardware or software sources such as keyboards, automatic speech recognition systems, cell phones, smart phones or the web onto an interpretation consisting of an action and one or more physical objects, such as robots, machinery, vehicles, etc. or digital objects such as data files, tables and databases. Tables and lists of (i) homonyms and misrecognitions, (ii) thematic relation patterns, and (iii) lexicons are used to generate alternative forms of the input which are scored to determine the best interpretation of the noisy input. The actions may be executed internally or output to any device which contains a digital component such as, but not limited to, a computer, a robot, a cell phone, a smart phone or the web. This invention may be implemented on sequential and parallel compute engines and systems.