Dynamic-Domain Speech Recognition With a Continuously Updated Entity Index
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems struggle with domain mismatch errors when processing voice queries in dynamic domains, such as the entertainment domain, due to their reliance on pre-trained data and inability to adapt to rapidly evolving content.
Innovation Solution
A two-level ASR system with a domain-specific entity index that is continuously updated with new entities, allowing for real-time adaptation and accurate processing of voice queries in dynamic domains, using a domain adapted audio command processing module that includes a grapheme-phoneme converter and fuzzy candidate generator.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional ASR systems use pre-trained data from public domains, then they can handle general speech inputs, but they cannot adapt to dynamic domains with new content
Solution Approach 1:
The system employs a two-level architecture where the first level uses a pre-trained ASR engine for general speech recognition, and the second level dynamically adapts to domain-specific terminology. The domain adaptation module continuously updates its knowledge base with new entities from dynamic domains, enabling the system to adapt to changing content without retraining the entire ASR model.
Solution Approach 2:
The ASR system is divided into two independent levels: a general-purpose ASR engine and a domain-specific adaptation layer. This segmentation allows the general model to remain stable while the adaptation layer handles domain-specific updates, resolving the contradiction between maintaining general capability and adapting to dynamic domains.
2Adaptability or versatility
If ASR systems are trained on broad speech domains, then they can handle multiple applications, but they make domain mismatch errors in specific dynamic domains
Solution Approach 1:
The system applies local quality by maintaining a general ASR model for broad coverage while adding domain-specific customization at the adaptation layer. Each dynamic domain receives tailored entity mappings and phonetic representations, improving local accuracy without compromising overall versatility.
Solution Approach 2:
The domain adaptation module acts as an intermediary between the general ASR engine and the dynamic domain requirements. It translates domain-specific entities into phonetic representations that the general ASR model can process, bridging the gap between broad coverage and specific accuracy.
3Reliability
If ASR systems are pre-configured with fixed lexicons, then they can process known entities, but they cannot recognize new or evolving terminology
Solution Approach 1:
The domain adaptation module operates continuously, continuously monitoring and updating its knowledge base with new entities from dynamic domains. This continuous action ensures the system always has the latest entity mappings available without interrupting the ASR processing flow.
Solution Approach 2:
The system performs preliminary action by pre-processing and indexing new entities into the domain adaptation module before they are encountered during ASR processing. This advance preparation ensures new terminology is immediately recognizable when it appears in speech inputs.
4Adaptability or versatility
If ASR systems are retrained with new data, then they can adapt to new domains, but the process is time-consuming and inefficient
Solution Approach 1:
Instead of retraining the entire ASR model, the system creates a lightweight copy or extension in the form of a domain adaptation module. This module copies the essential ASR functionality while adding domain-specific knowledge, enabling rapid adaptation without the time cost of full retraining.
Solution Approach 2:
The system applies partial action by updating only the domain adaptation module rather than the entire ASR system. This selective updating approach achieves necessary adaptation with minimal computational resources and time investment.
Data Source
AI summary
Disclosed herein are system, apparatus, article of manufacture, method, and computer program product embodiments for adapting an automated speech recognition system to provide more accurate suggestions to voice queries involving media content including recently created or recently available content. An example computer-implemented method includes transcribing the voice query, identifying respective components of the query such as the media content being requested and the action to be performed, and generating fuzzy candidates that potentially match the media content based on phonetic representations of the identified components. Phonetic representations of domain specific candidates are stored in a domain entities index and is continuously updated with new entries so as to maintain the accuracy of the speech recognition of voice queries for recently created or recently available content.