Sub-lexical Language Model for ASR Morphological Parsing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) systems face challenges in achieving accurate results in morphologically rich languages due to high out-of-vocabulary rates and data sparseness, particularly when operating with limited resources, as they struggle to model pronunciations for sub-lexical units and handle ambiguities in morphological parsing.
Innovation Solution
The approach involves segmenting words into sub-lexical units using linguistic or statistical methods, training a sub-lexical language model, and constructing a speech decoding network by composing a word to sub-lexical unit mapping transducer with a context dependency model and word level pronunciation lexicon, allowing for improved recognition accuracy by leveraging both word and sub-lexical level operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If word level language models are used for morphologically rich languages, then the system is simpler to implement, but the out-of-vocabulary rate increases and recognition accuracy decreases
Solution Approach 1:
The patent segments words into sub-lexical units (morphemes, roots, affixes) to create a hierarchical model. This segmentation allows the system to handle morphologically rich languages by breaking down complex words into manageable components, reducing the out-of-vocabulary rate while maintaining manageable system complexity through structured organization of the segmentation process.
Solution Approach 2:
The patent implements a nested hierarchical structure where sub-lexical units are nested within words, and words are nested within sentences. This nested doll approach allows the system to operate at multiple levels of abstraction simultaneously, combining the benefits of simple word-level processing with the flexibility of sub-lexical analysis for improved recognition accuracy.
2Reliability
If sub-lexical language models are used, then the out-of-vocabulary problem is reduced, but the system complexity and computational resources required increase
Solution Approach 1:
The system segments the language model into sub-lexical components that can be independently processed and combined. This segmentation reduces the overall complexity by allowing modular processing of morphemes and affixes rather than requiring the entire vocabulary to be processed as single units, making the sub-lexical approach more computationally feasible.
Solution Approach 2:
The patent merges word-level and sub-lexical level operations into a unified hierarchical model. By combining these approaches, the system leverages the simplicity of word-level processing for common words while utilizing sub-lexical analysis for complex or out-of-vocabulary words, optimizing the balance between accuracy and complexity.
3Reliability
If more resources are allocated to improve ASR accuracy, then the recognition results improve, but the computational cost and processing power requirements increase
Solution Approach 1:
The segmentation of language modeling into sub-lexical units enables more efficient resource utilization by processing only the necessary components of each word. This allows the system to achieve higher accuracy without proportionally increasing computational resources, as common morphemes and affixes can be processed more efficiently than complete words.
Solution Approach 2:
By merging word-level and sub-lexical operations, the system optimizes resource allocation by applying the most efficient processing level for each situation. This hybrid approach reduces overall computational cost while maintaining high recognition accuracy by leveraging the strengths of both methodologies.
Data Source
AI summary
An automatic speech recognition (ASR) system and method are provided for using sub-lexical language models together with word level pronunciation lexicons. These approaches operate by introducing a transduction between sequences of sub-lexical units and sequences of words.