Skill-Specific Word Embeddings for ASR Error Tolerance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face errors in natural language understanding due to inaccuracies in automatic speech recognition (ASR) processing, which can lead to misinterpretation of user inputs, especially when ASR fails to transcribe spoken words correctly.
Innovation Solution
The system generates skill-specific word embeddings using both text-based and ASR output data to improve the accuracy of natural language processing by adapting word vectors to reflect how users interact with the system, capturing variations and generalizations of user inputs, thereby reducing errors in natural language understanding (NLU) processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard word embeddings are used for NLU processing, then the system can process natural language inputs, but accuracy decreases when ASR transcription errors occur
Solution Approach 1:
The patent changes the parameters of word embeddings by generating skill-specific embeddings tailored to different domains (e.g., music, weather, news) rather than using generic embeddings. This allows the NLU system to adapt to domain-specific language patterns and be more tolerant of ASR errors within those domains, thereby improving both accuracy and robustness simultaneously
Solution Approach 2:
The patent segments the general NLU processing into skill-specific processing pathways. By dividing the monolithic embedding approach into multiple domain-specific embedding models, the system can handle different types of inputs with specialized representations, improving overall reliability across diverse scenarios
2Measurement precision
If skill-specific word embeddings are generated and used, then NLU processing accuracy improves, but system complexity increases
Solution Approach 1:
The patent creates a universal framework where skill-specific embeddings are generated on-demand rather than maintaining separate static models for each skill. The same embedding generation mechanism serves multiple skills by adapting to different domains, reducing overall system complexity while maintaining high accuracy
Solution Approach 2:
The system performs preliminary embedding generation by pre-processing text data and generating skill-specific embeddings in advance. This preliminary action allows the runtime system to simply select and apply pre-computed embeddings rather than performing complex computations in real-time, managing complexity effectively
Data Source
AI summary
The present disclosure describes the generation and use of word embeddings as part of natural language understanding (NLU) processing performed by a natural language processing system. In at least some examples, the word embeddings may be generated from text corpuses including at least text (representing spoken user inputs) output from automatic speech recognition (ASR) processing. In at least some examples, the word embeddings may be generated from text output from ASR processing and natural language text corresponding to one or more Internet webpages.