Speech Recognition Using Primitive Word Morphemes and Spacing Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face inefficiencies due to limited vocabulary in pronunciation dictionaries, leading to unrecognized words and increased resource usage, as they rely on predefined phoneme sequences, resulting in reduced recognition accuracy and increased resource consumption.
Innovation Solution
The implementation of a speech recognition apparatus and method that utilizes a processor to recognize speech using an acoustic model, a pronunciation dictionary, and a language model comprising primitive words, with a word-spacing model for correcting word spacing errors, where the dictionary and language model include words based on frequency, morpheme, or syllable-level criteria, and are trained on reconstructed corpora using techniques like recurrent neural networks or decision trees.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the number of words in the pronunciation dictionary is increased to recognize more words, then the recognition accuracy improves, but the resource usage increases dramatically
Solution Approach 1:
The patent segments the pronunciation dictionary into two parts: a limited-size dictionary containing only primitive words (high-frequency, basic vocabulary) and a separate OOV handling mechanism. This segmentation allows the system to maintain a small, efficient dictionary while still recognizing unlimited words through morpheme composition, resolving the contradiction between dictionary size and recognition accuracy.
Solution Approach 2:
The patent introduces morpheme-based intermediate units that serve as building blocks for constructing unrecognized words. Instead of storing every possible word, the system uses primitive words and morphemes as intermediaries to represent and recognize complex or unseen words through composition, reducing dictionary size while maintaining recognition capability.
2Quantity of substance
If the number of words in the pronunciation dictionary is limited to reduce resource usage, then the resource consumption decreases, but words not defined in the dictionary cannot be recognized
Solution Approach 1:
The patent implements a dynamic word construction mechanism where unrecognized words are assembled from primitive words and morphemes at runtime. Instead of a static dictionary containing all possible words, the system dynamically generates representations for OOV words using compositional morphology, enabling recognition of unlimited vocabulary with limited storage resources.
Solution Approach 2:
The patent performs preliminary extraction and storage of primitive words and morphemes during the training phase. By pre-identifying and storing only the essential building blocks (primitive words) rather than complete word forms, the system prepares for efficient OOV handling during recognition without requiring large dictionary storage.
3Device complexity
If traditional speech recognition methods are used with limited vocabulary, then the processing complexity remains manageable, but recognition rates decrease for unseen words
Solution Approach 1:
The patent replaces the traditional mechanical dictionary-lookup mechanism with a morpheme-based compositional system. Instead of relying on exact string matching against a limited dictionary, the system substitutes a linguistic analysis approach that decomposes words into morphemes and reconstructs meaning, enabling recognition of unseen words without proportionally increasing processing complexity.
Data Source
AI summary
A speech recognition apparatus includes a processor configured to recognize a user's speech using any one or combination of two or more of an acoustic model, a pronunciation dictionary including primitive words, and a language model including primitive words; and correct word spacing in a result of speech recognition based on a word-spacing model.