Garbage Model Phoneme Sub-word Selection for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional garbage models for speech recognition systems face challenges in balancing accuracy and processing efficiency, as they often include all possible phoneme combinations, leading to a large search space that decreases recognition speed and increases false alarms.
Innovation Solution
A garbage model is generated using a corpus of transcribed speech to constrain the search space by eliminating infrequently used phoneme combinations, incorporating phoneme transitions and speaker co-articulation information, and selecting a subset of sub-words based on quality and occurrence probabilities to improve accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If all possible phoneme combinations are included in the garbage model, then coverage of unforeseen speech is improved, but search space increases leading to decreased recognition speed
Solution Approach 1:
The patent extracts only the most frequently used phoneme combinations from the complete set of possible phoneme combinations. By taking out and retaining only the essential sub-words that cover the majority of unforeseen speech cases, the search space is dramatically reduced while maintaining effective coverage, thus resolving the contradiction between comprehensive coverage and recognition speed.
Solution Approach 2:
The patent changes the parameter of phoneme combination length by focusing on sub-words of specific lengths (e.g., 2-4 phonemes) rather than considering all possible lengths. This parameter change allows the system to capture the most common unforeseen speech patterns while limiting the search space to a manageable size, thereby improving recognition speed without sacrificing coverage.
2Adaptability or versatility
If all possible phoneme combinations are included in the garbage model, then coverage of unforeseen speech is improved, but false alarms increase
Solution Approach 1:
The patent extracts and retains only the most frequently occurring phoneme combinations from the complete set. By removing rare and infrequent combinations that are likely to cause false alarms, the system maintains coverage of common unforeseen speech while significantly reducing the harmful effect of false positive recognitions.
Solution Approach 2:
The patent applies partial action by considering only a subset of phoneme combinations (those with frequency above a threshold) rather than all possible combinations. This partial approach is sufficient to cover the majority of meaningful unforeseen speech cases while avoiding the excessive action of including rare combinations that would generate false alarms.
3Measurement precision
If a large number of sub-words are included in the garbage model, then accuracy of unforeseen speech recognition is improved, but processing complexity increases
Solution Approach 1:
The patent changes the parameter of sub-word frequency by introducing a frequency threshold. Only sub-words with occurrence frequency above this threshold are included in the garbage model. This parameter change ensures that the model includes enough sub-words to maintain high accuracy for unforeseen speech recognition, while excluding low-frequency sub-words that would increase processing complexity without providing proportional accuracy benefits.
Data Source
AI summary
Methods and apparatus for performing speech recognition using a garbage model. The method comprises receiving audio comprising speech and processing at least some of the speech using a garbage model to produce a garbage speech recognition result. The garbage model includes a plurality of sub-words, each of which corresponds to a possible combination of phonemes in a particular language.