Speech Recognition Corpus Update via Search Query Frequency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic Speech Recognition (ASR) systems face challenges in accurately transcribing spoken utterances, especially when the acoustic model is not trained for the specific domain or dialect, leading to reduced accuracy and limited adaptability to different speakers.
Innovation Solution
The method involves receiving information on the frequency of search queries submitted to a search engine and determining groupings of words based on their order in the query sequence. These groupings are then provided to an ASR system to update its corpus of sequences, allowing it to improve transcription accuracy by focusing on popular or spiking queries, which are represented in a factor graph that constrains the search space and assigns higher probabilities to correct word sequences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the ASR system uses a comprehensive corpus of all possible word sequences, then transcription accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent extracts only the most relevant and frequently occurring word sequences from the comprehensive corpus based on search query frequency. By identifying and isolating high-frequency n-grams from the vast search query data, the system maintains transcription accuracy for common phrases while reducing the overall corpus size and computational burden.
Solution Approach 2:
The patent applies different quality levels to different parts of the corpus by prioritizing high-frequency search queries over low-frequency ones. The system creates a tiered corpus structure where frequently searched terms receive higher priority and more processing resources, while less common terms use default or lower-priority handling, optimizing the balance between accuracy and computational efficiency.
2Measurement precision
If the ASR system adapts to domain-specific terminology and dialects, then transcription accuracy for specific domains is improved, but the system's versatility across different domains may be reduced
Solution Approach 1:
The patent implements a dynamic corpus that can be updated and reweighted based on changing search query patterns. The system continuously monitors search query frequencies and adjusts the corpus composition accordingly, allowing it to adapt to emerging domains and terminology while maintaining its core functionality across established domains. This dynamic updating mechanism enables the system to remain versatile while specializing in hot topics.
Solution Approach 2:
The patent creates a multi-functional corpus structure that serves both general-purpose transcription and domain-specific optimization. By organizing the corpus to include both common vocabulary and domain-specific terminology from search queries, the system can handle a wide variety of transcription tasks across different domains while maintaining high accuracy in specialized areas.
3Quantity of substance
If the ASR system processes all search queries equally, then comprehensive coverage is improved, but processing efficiency and speed decrease
Solution Approach 1:
The patent applies partial action by focusing computational resources on processing only the most frequent and relevant search queries in detail, while handling less frequent queries with more streamlined or approximate methods. This selective processing approach maintains comprehensive coverage over time while significantly improving processing efficiency for the majority of common queries.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods and systems for speech recognition processing are described. In an example, a computing device may be configured to receive information indicative of a frequency of submission of a search query to a search engine for a search query composed of a sequence of words. Based on the frequency of submission of the search query exceeding a threshold, the computing device may be configured to determine groupings of one or more words of the search query based on an order in which the one or more words occur in the sequence of words of the search query. Further, the computing device may be configured to provide information indicating the groupings to a speech recognition system.