Automatic ASR Dictionary Learning via Phoneme Tokenization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems require manual updates and significant developer effort to recognize new words, entities, and pronunciations, leading to a frustrating user experience due to their inability to automatically learn and adapt to evolving language usage.
Innovation Solution
The technology enables automatic learning of entities, words, pronunciations, and parts of speech by recognizing phoneme sequences, tokenizing them, and adding new tokens or pronunciations to dictionaries based on occurrence counts and edit distances, allowing ASR systems to dynamically update and improve without manual intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual updates are used to add new words and entities to the system, then the pronunciation dictionary and entity lists remain accurate and controlled, but the system cannot keep up with the fast pace of new word entry and requires large amounts of developer effort
Solution Approach 1:
The system automatically learns new words, pronunciations, entities, and parts of speech from speech audio data without requiring manual developer intervention. The ASR system performs self-updating by detecting unknown phoneme subsequences, verifying their validity through occurrence counting, and automatically adding them to the pronunciation dictionary and entity lists, thereby serving itself rather than requiring external manual updates
Solution Approach 2:
The system performs preliminary verification and validation steps before adding new words to the dictionary. It counts occurrences of unknown phoneme subsequences across multiple speech audio segments, checks if they meet threshold criteria, and validates them through automated processes before incorporation, ensuring accuracy while maintaining speed
2Productivity
If the system automatically adds every unrecognized phoneme sequence to the dictionary, then new words are quickly captured, but incorrect or rare words may be falsely added
Solution Approach 1:
The system implements feedback mechanisms by counting the occurrences of unknown phoneme subsequences across multiple speech audio segments. It uses this feedback information to determine whether a phoneme subsequence should be added to the dictionary, only incorporating those that meet predetermined occurrence thresholds and statistical criteria, thereby filtering out false positives while capturing genuine new words
Solution Approach 2:
The system takes a conservative approach by requiring multiple occurrences and meeting threshold criteria before adding a new word, rather than adding every single unrecognized sequence. This partial action approach ensures higher reliability by only incorporating words that demonstrate sufficient evidence of being legitimate new entries
3Adaptability or versatility
If a large pronunciation dictionary covering all possible words is maintained, then the system can recognize any word, but the system complexity and storage requirements increase significantly
Solution Approach 1:
The pronunciation dictionary is implemented as a dynamic, evolving data structure that automatically grows and adapts based on actual usage patterns. Rather than maintaining a static, pre-completed dictionary, the system continuously learns from speech audio data and updates the dictionary in real-time, allowing it to achieve comprehensive coverage while keeping the initial dictionary size manageable
Solution Approach 2:
The system segments the vocabulary into known dictionary words and unknown phoneme subsequences. By processing and analyzing unknown subsequences separately through automated learning, the system can efficiently expand its vocabulary without requiring the entire dictionary to be pre-populated, thereby reducing initial complexity while maintaining adaptability
Data Source
AI summary
Systems for automatic speech recognition and/or natural language understanding automatically learn new words by finding subsequences of phonemes that, if they were a new word, would enable a successful tokenization of a phoneme sequence. Systems can learn alternate pronunciations of words by finding phoneme sequences with a small edit distance to existing pronunciations. Systems can learn the part of speech of words by finding part-of-speech variations that would enable parses by syntactic grammars. Systems can learn what types of entities a word describes by finding sentences that could be parsed by a semantic grammar but for the words not being on an entity list.