Transcription Knowledge Graph for Dynamic ASR Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition systems struggle with correcting mis-transcriptions in dynamic domains, such as the entertainment domain, due to their pre-training on static data and inability to adapt to rapidly evolving content.
Innovation Solution
A transcription knowledge graph system that utilizes a voice graph and phoneme-embedding module to correct ASR errors by training a machine learning embedding model to generate numeric representations of transcriptions, allowing for the identification of similar entities even in dynamic domains without requiring human intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional ASR systems use pre-training on static data, then the system structure remains simple and stable, but the system cannot adapt to rapidly evolving content in dynamic domains
Solution Approach 1:
The patent introduces a transcription knowledge graph as an intermediary component between the ASR engine and the final transcription output. This knowledge graph stores domain-specific entities, relationships, and contextual information, enabling the system to adapt to dynamic domains without modifying the underlying ASR engine. The knowledge graph acts as a mediator that enriches ASR outputs with domain knowledge, resolving the contradiction between maintaining simple ASR system structure and achieving adaptability to evolving content.
Solution Approach 2:
The system segments the transcription process into distinct components: the ASR engine for speech-to-text conversion, the transcription knowledge graph for domain-specific knowledge storage, and the entity resolution module for matching and disambiguation. This segmentation allows each component to remain relatively simple while the integrated system achieves high adaptability. The knowledge graph can be independently updated with new domain information without affecting the ASR engine.
2Measurement precision
If the system implements real-time adaptation to new entities, then the accuracy of transcription correction improves, but the processing time and computational resources increase
Solution Approach 1:
The system performs preliminary actions by pre-building and storing domain-specific knowledge in the transcription knowledge graph before real-time transcription processing. Entities, relationships, and contextual information are organized and indexed in advance, enabling rapid retrieval during real-time operation. This preliminary preparation allows the system to achieve high transcription accuracy without incurring excessive processing time during actual use, as the heavy computational work of organizing knowledge is done beforehand.
Solution Approach 2:
The system applies partial action by selectively processing only the portions of transcription that require correction or disambiguation, rather than re-processing the entire transcription. The entity resolution module focuses computational resources on identifying and correcting specific errors by querying the knowledge graph for relevant entities and relationships, thereby achieving high accuracy while minimizing overall processing time.
3Measurement precision
If the system uses a voice graph with n-best outputs to correct ASR errors, then the transcription accuracy improves, but the device complexity increases
Solution Approach 1:
The system extracts only the essential elements needed for error correction from the n-best ASR outputs and stores them in the transcription knowledge graph. Rather than maintaining complex relationships among all possible transcriptions, the knowledge graph extracts and stores only the most probable correct transcriptions along with their contextual information and entity associations. This extraction approach maintains high error correction accuracy while simplifying the overall system structure by focusing on critical information only.
Data Source
AI summary
Some embodiments include a transcription knowledge graph that can resolve automatic speech recognition (ASR) engine output errors. In some embodiments, a transcription knowledge graph can utilize data from past sessions of the ASR engine to form a voice graph that can be analyzed to determine a correlation between a mis-transcription (error text) and the correct transcription (correct text). Thus, ASR engine outputs, even if they include a mis-transcription, can be adjusted to the correct transcription. Further, the correct transcriptions and the voice graph can be used to train machine learning (ML) algorithms to generate numerical representations of an entity. The ML algorithms can be applied to a transcription to correctly identify a corresponding entity label, even if the transcription was not utilized in the voice graph to train the ML algorithm.