Prefix Tree Bonus Scores for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition models, particularly end-to-end STT systems, face challenges in accurately recognizing new words that are not adequately represented in their prefix trees, leading to suboptimal performance when encountering rare or untrained vocabulary.
Innovation Solution
A method is introduced where a bonus score is assigned to transitions in the prefix tree for new words, enhancing the probability of selecting these words in the recognized text sequence. This involves adding new words to the prefix tree, assigning bonus scores to relevant transitions, and adjusting scores based on the presence or absence of these transitions in the speech sequence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If new words are added to the prefix tree for speech recognition, then the recognition accuracy of rare or untrained vocabulary is improved, but the complexity of the prefix tree structure increases
Solution Approach 1:
The patent applies local quality by assigning bonus scores to specific transitions in the prefix tree that correspond to new or rare words, rather than uniformly modifying the entire tree structure. This localized modification allows the system to improve recognition accuracy for specific vocabulary while maintaining the overall simplicity of the prefix tree structure.
Solution Approach 2:
The patent changes the scoring parameter of the prefix tree by introducing bonus scores that are added to transitions representing new words. This parameter modification enables the system to recognize new vocabulary without requiring extensive retraining or structural changes to the prefix tree, thus improving recognition accuracy while controlling complexity.
2Reliability
If bonus scores are assigned to transitions representing new words, then the probability of selecting new words in output text is improved, but the computational complexity of score calculation increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing bonus scores for transitions in the prefix tree before speech recognition occurs. This allows the system to quickly retrieve and add these pre-computed scores during recognition, improving the reliability of new word selection without significantly increasing real-time computational complexity.
Solution Approach 2:
The system performs self-service by automatically managing the bonus score assignments and hypothesis score calculations without requiring external intervention or extensive retraining. The prefix tree structure itself facilitates the score calculation by organizing transitions in a way that enables efficient computation of hypothesis scores incorporating bonus scores.
3Loss of time
If the prefix tree is modified to include new words without retraining the model, then the training time and computational resources are reduced, but the integration of new vocabulary into the recognition system becomes simpler
Solution Approach 1:
The patent applies segmentation by separating the vocabulary update process from the model training process. New words are added to the prefix tree as independent transitions with assigned bonus scores, rather than requiring complete retraining of the speech recognition model. This segmentation allows for quick vocabulary updates while maintaining the existing trained model, thus reducing training time and simplifying vocabulary integration.
Data Source
AI summary
An approach for improving speech recognition is provided. A processor receives a new word to add to a prefix tree. A processor determines a bonus score for a first transition from a first node to a second node in a prefix tree on condition that the first transition is included in a path of at least one transition representing the new word. A processor determines a hypothesis score for a hypothesis that corresponds to a speech sequence based on the prefix tree, where the hypothesis score adds the bonus score to an initial hypothesis score to determine the hypothesis score. In response to a determination that the hypothesis score exceeds a threshold value, a processor generates an output text sequence for the speech sequence based on the hypothesis.