ASR Saliency Weighting for User-Specific Transcription Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic Speech Recognition (ASR) technologies face challenges in accurately measuring accuracy, as they consider all errors equally, failing to account for the varying importance of words to different users and the context of messages, leading to suboptimal transcription results.
Innovation Solution
Assigning saliency weights to words within ASR models based on human perception judgments, which are stored in user profiles and applied to weight transcripts, allowing for context-dependent weighting schemes such as linear, logarithmic, or inverse weighting, enabling more accurate and user-specific transcription.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional Word Error Rate (WER) is used to measure ASR accuracy, then all errors are treated equally, but this fails to capture the varying importance of different words to users
Solution Approach 1:
The patent applies local quality by assigning different saliency weights to different words based on their importance to the user. Instead of treating all words equally, the system identifies salient words (such as nouns, proper nouns, or context-specific terms) and assigns them higher weights, while non-salient words receive lower weights. This allows the accuracy measurement to reflect the varying importance of different words to the user's understanding of the message.
Solution Approach 2:
The patent implements dynamics by making the saliency weights adaptive and context-dependent. The system can adjust word weights based on the specific conversation context, user preferences, and message type. This dynamic weighting allows the ASR system to adapt to different users and situations, improving the relevance and accuracy of the transcription for each specific use case.
2Ease of operation
If fixed saliency weights are assigned to words, then processing is simplified, but the system cannot adapt to individual user needs and varying message contexts
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing saliency weights for words in a database before actual ASR processing. The system maintains a dictionary of salient words with pre-determined weights that can be quickly retrieved during transcription. This preliminary preparation simplifies the real-time processing while still enabling user-specific customization, as the pre-computed weights can be selected based on the user profile and message context without adding computational complexity during the actual transcription process.
3Ease of manufacture
If all words are transcribed with equal importance, then the ASR system is easier to implement, but the transcription does not capture the meaningful content that users care about
Solution Approach 1:
The patent applies the extraction principle by identifying and separating salient words from non-salient words in the transcription process. The system extracts only the most important words (such as key nouns, proper nouns, and context-critical terms) and applies enhanced weighting to these extracted elements. This allows the ASR system to focus computational resources on capturing the most meaningful content while maintaining simpler processing for less important words, thereby reducing information loss of critical content without requiring complete redesign of the entire system.
Data Source
AI summary
Disclosed herein are systems, methods, and non-transitory computer-readable storage media for assigning saliency weights to words of an ASR model. The saliency values assigned to words within an ASR model are based on human perception judgments of previous transcripts. These saliency values are applied as weights to modify an ASR model such that the results of the weighted ASR model in converting a spoken document to a transcript provide a more accurate and useful transcription to the user.