Dynamic-Domain Speech Recognition Index for Media Query Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems struggle with domain mismatch errors when processing voice queries involving dynamic domains like entertainment content, where new words and pronunciations constantly evolve, leading to inaccurate translations.
Innovation Solution
A two-level ASR system with a domain-specific entity index that is continuously updated with new entities, using phonetic representations to generate fuzzy candidates that accurately match dynamic domain queries, without requiring retraining of the ASR engine.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If ASR systems are trained on large amounts of public domain data, then they can handle a broad range of speech domains, but they cannot adapt to new dynamic domains without retraining
Solution Approach 1:
The system divides the ASR functionality into two independent parts: a pre-trained general ASR engine and a domain-specific entity index. This segmentation allows the entity index to be updated independently without retraining the ASR engine, enabling rapid adaptation to new domains like entertainment content while maintaining the original model's performance.
Solution Approach 2:
The system introduces a dynamic entity index that can be continuously updated with new entertainment entities, words, and pronunciations. This dynamic structure allows the system to adapt to changing domains in real-time without requiring retraining of the underlying ASR engine, directly addressing the adaptability challenge.
2Adaptability or versatility
If ASR systems use generalist training for broad application, then they can handle multiple speech domains, but they produce domain mismatch errors in dynamic domains
Solution Approach 1:
The system applies local quality by maintaining a domain-specific entity index tailored to entertainment content, which provides specialized knowledge for that domain. This local optimization complements the general ASR engine, improving recognition accuracy for entertainment entities while preserving the ability to handle other domains through the general model.
Solution Approach 2:
The domain-specific entity index acts as an intermediary between the general ASR engine and the dynamic entertainment domain. It receives transcriptions from the ASR engine and provides corrected entity mappings, mediating the interaction between the general model and domain-specific requirements to eliminate domain mismatch errors.
3Measurement precision
If ASR systems are retrained with new domain data, then they can improve accuracy in that domain, but the process is computationally expensive and time-consuming
Solution Approach 1:
The system extracts the domain-specific adaptation requirements from the main ASR training process and places them in a separate, manageable entity index. This extraction allows domain-specific improvements to be achieved without retraining the entire ASR engine, significantly reducing computational complexity while maintaining accuracy.
Solution Approach 2:
Instead of copying and retraining the entire ASR engine for each new domain, the system creates a lightweight copy of the general model and supplements it with a domain-specific entity index. This copying approach enables rapid domain adaptation with minimal computational resources required.
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.