Transcription Knowledge Graph with Phoneme Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems struggle with dynamic domains such as the entertainment domain, where new words and pronunciations constantly evolve, leading to mis-transcriptions that cause malfunctions and negative user experiences due to their pre-training on static data and difficulty in updating.
Innovation Solution
A transcription knowledge graph system that includes a voice graph ASR error correction module and a phoneme-embedding module to correct ASR output errors by training a machine learning embedding model to generate numeric representations of phonetic entities, allowing for the identification of correct entities even in dynamic domains without requiring human intervention or updating the ASR engine.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Stability of the object's composition
If conventional ASR systems use pre-trained static data models, then system stability is maintained, but adaptability to dynamic domains deteriorates
Solution Approach 1:
The patent introduces a phoneme-embedding model as an intermediary layer between the ASR engine and the transcription output. This mediator captures phonetic representations of transcriptions and enables adaptation to new entities without retraining the entire ASR system, thus maintaining system stability while improving adaptability to dynamic domains
Solution Approach 2:
The patent segments the ASR system into distinct components: the pre-trained ASR engine and the separate phoneme-embedding model. This segmentation allows the phoneme-embedding model to be independently trained and updated with new entities while the main ASR engine remains stable and unchanged
2Adaptability or versatility
If ASR engine is updated with new data, then adaptability to new entities improves, but implementation complexity and cost increase
Solution Approach 1:
The patent extracts the phonetic representation capability from the complex ASR engine into a separate, simpler phoneme-embedding model. This extracted component can be trained independently on new entities using relatively simple training data (transcriptions and their phoneme embeddings), avoiding the complexity of updating the entire ASR engine
3Measurement precision
If conventional ASR systems rely on exact transcription matching, then precision in entity identification is maintained, but coverage of ASR errors deteriorates
Solution Approach 1:
The patent changes the parameter used for entity identification from exact transcription text matching to phoneme-embedding vector similarity. This parameter change allows the system to identify entities based on phonetic similarity rather than exact text match, thereby covering a broader range of ASR errors including mis-transcriptions while maintaining precision through vector similarity comparison
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.