Speech Recognition Context Modeling for Long-Tail Word Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies suffer from low accuracy in recognizing long-tail words due to insufficient training on these less frequently used words, leading to inadequate contextual information.
Innovation Solution
A speech recognition method that obtains a first speech and text, constructs a set with text identifications and features associated with subsequent texts, and determines text content based on this set to improve prediction accuracy by incorporating more long-tail word information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech recognition models are used, then the recognition process is simple, but the accuracy for long-tail words is low
Solution Approach 1:
The patent embeds a text feature extraction module within the speech recognition model. This nested structure allows the model to incorporate text features (including long-tail word information) as additional input features, enhancing recognition accuracy without requiring a complete redesign of the speech recognition architecture
Solution Approach 2:
The patent performs preliminary extraction of text features from historical speech data before the actual speech recognition process. By pre-processing and storing text features (especially for long-tail words) in advance, the model can directly utilize these features during recognition, improving accuracy while avoiding complex real-time processing
2Measurement precision
If more training data is collected to improve long-tail word recognition, then the contextual information improves, but the data collection and processing time increases
Solution Approach 1:
The patent extracts and stores text features from historical speech data in advance, creating a pre-processed feature database. This preliminary action eliminates the need for real-time data collection and processing during speech recognition, significantly reducing time loss while maintaining improved accuracy for long-tail words
Solution Approach 2:
The patent creates feature representations (copies) of historical speech data in the form of text features. Instead of processing raw historical data during recognition, the system uses these pre-extracted feature copies, which contain essential information about long-tail words and contextual patterns, thereby avoiding time-consuming re-processing of original data
Data Source
AI summary
A speech recognition method and apparatus, and an electronic device. The method comprises: obtaining a first speech; obtaining a first text corresponding to a previous segment of speech of the first speech; obtaining a first set, the first set comprising a plurality of text identifications and a text feature corresponding to each of the plurality of text identifications, the text feature being a feature associated with a plurality of subsequent texts of a text corresponding to the text identification, the text feature being associated with frequencies of the plurality of subsequent texts of the text in a text set, and the first set being determined based on the text set; and determining, based on the first text and the first set, text content associated with the first speech.