Out-of-Vocabulary Search in Speech Recognition via Anchor Segments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic Speech Recognition (ASR) systems, particularly Large Vocabulary Continuous Speech Recognition (LVCSR), struggle to search for phrases containing Out-Of-Vocabulary (OOV) words due to their reliance on predefined vocabularies, leading to inaccurate recognition and inability to find instances of OOV words in audio searches.
Innovation Solution
A system and method that identifies anchor segments within audio files for OOV words in a LVCSR output, allowing for a modified recognition phase to generate new search events, enabling faster processing by focusing on specific, smaller parts of the audio corpus rather than reprocessing the entire corpus.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the OOV word is added to the dictionary and a new LM is generated, then the search accuracy for OOV words is improved, but the processing time increases due to reprocessing the audio corpus
Solution Approach 1:
The patent segments the audio corpus into smaller anchor segments based on acoustic similarity to the OOV query. Instead of reprocessing the entire audio corpus, the system identifies and processes only the relevant segments that are acoustically similar to the query, thereby maintaining search accuracy while significantly reducing processing time.
Solution Approach 2:
The system performs preliminary acoustic similarity matching to identify anchor segments before conducting the full speech recognition search. This preliminary action pre-screens the audio corpus to locate only those segments that are likely to contain the OOV query, avoiding the need to reprocess the entire corpus and reducing overall processing time.
2Adaptability or versatility
If phonetic transcription is used for OOV search, then the search flexibility is improved, but the recognition accuracy deteriorates
Solution Approach 1:
The patent segments the search process into two stages: first using phonetic transcription to identify anchor segments (providing flexibility), then applying full speech recognition with constrained grammar to those segments (providing accuracy). This segmentation allows the system to leverage the strengths of both approaches.
Solution Approach 2:
The patent uses anchor segments as an intermediary between the phonetic transcription search and the final recognition result. The phonetic search identifies candidate regions, and the constrained grammar search on these regions produces the final accurate results, with the anchor segments serving as the mediating structure that connects both approaches.
3Reliability
If the entire audio corpus is reprocessed to search for OOV words, then the search completeness is improved, but the processing speed deteriorates
Solution Approach 1:
The patent extracts only the relevant anchor segments from the entire audio corpus based on acoustic similarity to the OOV query. By taking out and processing only these extracted segments rather than the entire corpus, the system maintains search completeness for OOV words while dramatically improving processing speed.
Solution Approach 2:
The system performs partial reprocessing by applying constrained grammar search only to the identified anchor segments rather than the entire audio corpus. This partial action is sufficient to find all OOV instances while avoiding the excessive processing time that would result from reprocessing the complete corpus.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method including: receiving, on a computer system, a text search query, the query including one or more query words; generating, on the computer system, for each query word in the query, one or more anchor segments within a plurality of speech recognition processed audio files, the one or more anchor segments identifying possible locations containing the query word; post-processing, on the computer system, the one or more anchor segments, the post-processing including: expanding the one or more anchor segments; sorting the one or more anchor segments; and merging overlapping ones of the one or more anchor segments; and searching, on the computer system, the post-processed one or more anchor segments for instances of at least one of the one or more query words using a constrained grammar.