Topic Model Query Expansion for Speech Recognition Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large Vocabulary Continuous Speech Recognition (LVCSR) systems suffer from high word error rates (WER) of 30%-50%, leading to reduced recall of searched terms in transcripts, necessitating a method to enhance search query expansion.
Innovation Solution
A method that generates a topic model from LVCSR transcripts, selects keywords for candidate term expansion based on phonetic and semantic similarities, and uses logical operators to expand search queries, improving retrieval precision and recall by identifying substitution errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If LVCSR systems are used to generate transcripts, then vocal interactions can be converted to text for search, but high word error rates (30%-50%) significantly reduce the recall of searched terms
Solution Approach 1:
The patent introduces topic models as an intermediary between LVCSR transcripts and search queries. The topic model comprises multiple topics, each with associated keywords and phonetic transcriptions. When a search term is queried, the system identifies relevant topics and expands the search to include phonetically similar keywords from those topics, thereby mediating the high error rate of LVCSR and improving term recall without sacrificing transcript generation efficiency
Solution Approach 2:
The system changes the search parameters by expanding single-term queries into multi-term searches based on phonetic similarities. Instead of searching for exactly one term, the system transforms the search to include multiple candidate terms derived from topic model keywords, effectively changing the search parameter from precise matching to phonetic similarity matching, which compensates for LVCSR errors
2Measurement precision
If search queries are expanded with multiple candidate terms, then recall of searched terms improves, but search complexity and processing time increase
Solution Approach 1:
The topic model is pre-computed and stored before actual search operations. During search, the system only needs to retrieve pre-computed keywords and phonetic transcriptions from the topic model rather than performing complex analysis in real-time. This preliminary action reduces search-time complexity while maintaining improved recall through expanded candidate terms
3Measurement precision
If phonetic transcriptions are extracted and compared for term expansion, then substitution errors can be identified, but processing time and computational resources increase
Solution Approach 1:
The system creates phonetic transcriptions (copies of phonetic information) for both search terms and topic model keywords. Instead of performing complex phonetic analysis during search, the system compares these pre-extracted phonetic copies directly. This copying approach enables efficient substitution error detection by comparing phonetic representations rather than analyzing raw audio or text during the search process
Data Source
AI summary
The subject matter discloses a method for expansion of search queries on large vocabulary continuous speech recognition transcripts comprising: obtaining a textual transcript of audio interaction generated by the large vocabulary continuous speech recognition; generating a topic model from the textual transcripts; said topic model comprises a plurality of topics wherein each topic of the plurality of topics comprises a list of keywords; obtaining a search term; associating a topic from the topic model with the search term; and generating a list of candidate term expansion words by selecting keywords from the list of keywords of the associated topic; said candidate term expansion words are of high probability to be substitution errors of the search term that are generated by the large vocabulary continuous speech recognition.