Context-Aware Voice Query Transcription via Language Model Biasing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional automated speech recognition systems often misrecognize voice queries due to their inability to distinguish between terms commonly used in different user contexts, leading to inaccurate transcriptions.
Innovation Solution
The method generates multiple transcription candidates for a voice query by using user context data, such as location information and previous search queries, to identify phonetically similar terms that are not initially recognized, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional automated speech recognition systems compare audio to a dictionary of words without context, then the system operation is simple and fast, but the recognition accuracy deteriorates due to inability to distinguish between phonetically similar terms in different contexts
Solution Approach 1:
The system performs preliminary actions by obtaining user context data (location, previous queries, user profile) before conducting speech recognition. This preliminary context acquisition enables the system to generate context-aware candidate transcriptions that are biased toward contextually relevant terms, thereby improving recognition accuracy without significantly increasing operational complexity during the actual speech processing phase
Solution Approach 2:
The speech recognition process is segmented into multiple stages: generating initial candidate transcriptions from audio, obtaining user context data, generating additional context-aware candidate transcriptions, and selecting the best match. This segmentation allows the system to apply different processing strategies at each stage, improving overall accuracy while managing complexity through modular processing
2Measurement precision
If the system generates multiple transcription candidates based on user context, then the recognition accuracy improves, but the processing time increases
Solution Approach 1:
The system applies partial action by generating additional candidate transcriptions selectively based on context relevance rather than exhaustively generating all possible transcriptions. The language model biases candidate generation toward contextually probable terms, reducing the search space and processing time while maintaining improved accuracy through context-aware candidate selection
3Reliability
If the system uses user context data to generate additional candidate transcriptions, then the likelihood of accurate transcription improves, but the computational resources required increase
Solution Approach 1:
The system changes parameters by adjusting the language model bias based on user context. Instead of performing computationally intensive full speech recognition multiple times, the system modifies the probability parameters of the language model to favor contextually relevant terms during candidate generation, achieving improved reliability with reduced computational energy consumption
Data Source
AI summary
Methods, including computer programs encoded on a computer storage medium, for collaborative language model biasing. In one aspect, a method includes: obtaining (i) one or more initial candidate transcriptions, and (ii) one or more terms that are associated with a context; selecting one or more of the terms that are associated with the context, and that (i) do not occur in the candidate transcriptions, and (ii) are indicated as phonetically similar to one or more terms that do occur in the initial candidate transcriptions; generating one or more additional candidate transcriptions based on the (i) initial candidate transcriptions, and (ii) the selected terms; and providing the one or more additional candidate transcriptions to an automated speech recognizer.