Phonetic Matching for ASR Named Entity Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition systems face high error rates when encountering fanciful or uncommon brand names, leading to inaccuracies in speech recognition, natural language understanding, and dialogue management, requiring retraining of models which can take weeks or months.
Innovation Solution
A novel named entity model is introduced within the speech processing component that converts text output from automatic speech recognition into phonetic representations, matching these with a database of named entities to correct errors, allowing for accurate recognition of brand names and other misrecognized words.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If automatic speech recognition modules are trained on standard words and normal conversations, then the system performs well on common dialogues, but it produces high error rates when encountering fanciful or new brand names
Solution Approach 1:
The system performs preliminary actions by converting speech recognition output to phonetic representations and comparing against a pre-built database of named entities before final interpretation. This preliminary phonetic matching allows the system to correct errors in real-time without requiring full retraining, thus maintaining reliability while improving adaptability to new brand names.
Solution Approach 2:
The patent introduces phonetic representations as an intermediary between speech recognition output and natural language understanding. This intermediary layer enables accurate matching of brand names by comparing phonetic patterns rather than relying solely on text-based recognition, resolving the contradiction between recognizing standard words and adapting to new brand names.
2Reliability
If the automatic speech recognition module is retrained to recognize new brand names, then recognition accuracy improves, but the process takes 2 to 3 months
Solution Approach 1:
The system extracts the specific problem of brand name recognition from the overall speech recognition system by creating a separate phonetic matching layer that operates independently. This extracted component can be updated and maintained separately without requiring full system retraining, reducing the time investment from months to a fraction of that time.
Solution Approach 2:
The patent creates phonetic copies or representations of brand names in a database that can be quickly added and referenced. Instead of retraining the entire speech recognition model on new brand names, the system copies the phonetic patterns into the existing database structure, enabling rapid adaptation without lengthy retraining processes.
3Measurement precision
If speech recognition errors occur on brand names, then the text output becomes inaccurate, but correcting these errors through retraining is inefficient
Solution Approach 1:
The system implements feedback by continuously comparing phonetic representations of recognized text against the named entity database and automatically correcting discrepancies. This real-time feedback mechanism improves text output accuracy without manual intervention or retraining, maintaining high productivity while achieving precise error correction.
Solution Approach 2:
The phonetic matching system performs self-service error correction by automatically identifying and correcting brand name recognition errors without requiring external retraining or manual correction. The system serves itself by using its own phonetic database to fix errors in real-time, improving accuracy while maintaining efficiency.
Data Source
AI summary
A new approach to automatic speech recognition is disclosed. An example method include receiving a first text representing speech recognition of a phrase spoken by a user, isolating a candidate named entity from within the phrase, receiving a first phonetic representation of the candidate named entity, comparing the first phonetic representation to phonetic representations in a mapping database which map the phonetic representations to words to yield a comparison, based on the comparison, identifying a second phonetic representation in the mapping database that matches a second text in the mapping database to the second phonetic representation and replacing the candidate named entity with the second text. The approach can be used for new brands for which automatic speech recognition error rates are high.