ASR Language Modeling for Non-Lexical Entities
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) systems face challenges in recognizing non-lexical entities, such as URLs, email addresses, and dollar amounts, due to data sparsity and high out-of-vocabulary rates, which complicates language modeling and pronunciation rendering.
Innovation Solution
The implementation of a written-domain language modeling approach using finite-state modeling techniques, including decomposition-recomposition methods and restriction models, to address the challenges of non-lexical entities, allowing for improved recognition and pronunciation of these entities by decomposing them into lexical units and ensuring proper boundaries are enforced.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional language modeling is used for non-lexical entities, then the ASR system can recognize regular words, but it fails to properly recognize non-lexical entities due to data sparsity and high out-of-vocabulary rates
Solution Approach 1:
The patent applies segmentation by decomposing non-lexical entities into smaller lexical units using rewrite grammar rules. For example, a URL like 'www.example.com' is segmented into 'www', 'example', 'com' components. This decomposition transforms sparse non-lexical entity data into abundant lexical unit data that can be effectively modeled by n-gram language models, thereby resolving the data sparsity problem while maintaining recognition accuracy.
2Adaptability or versatility
If non-lexical entities are treated as single units, then the vocabulary size increases, but the out-of-vocabulary rate increases and language modeling becomes more challenging
Solution Approach 1:
Instead of treating non-lexical entities as single complex vocabulary items, the patent segments them into smaller lexical units through rewrite rules. This segmentation reduces the out-of-vocabulary rate because the segmented components are more likely to be found in the training data, making language modeling more effective while maintaining the ability to handle diverse non-lexical entities.
Solution Approach 2:
The patent introduces a restriction model as an intermediary component that works between the acoustic model and the language model. This restriction model enforces proper boundaries for non-lexical entities by restricting language model paths, thereby simplifying the overall language modeling task while maintaining vocabulary coverage and recognition accuracy.
3Measurement precision
If decomposition rewrite rules are applied to training data, then non-lexical entities are properly segmented, but the processing complexity increases
Solution Approach 1:
The patent applies decomposition rewrite rules during the offline training data preparation phase, before actual speech recognition occurs. By performing the decomposition and segmentation work in advance during training, the system establishes proper boundaries and structures in the training data without adding complexity to the online recognition process. The restriction model is also pre-computed from the decomposed training data, further reducing online processing complexity.
Data Source
AI summary
An automatic speech recognition system and method are provided for written-domain language modeling. According to one implementation, a process includes accessing decomposed training data that results from applying rewrite grammar rules to original training data, the decomposed training data comprising (i) regular words from the original training data that have not been rewritten using the set of rewrite grammar rules, and (ii) decomposed segments that result from rewriting non-lexical entities from the original training data using the rewrite grammar rules, generating a restriction model that (i) maps language model paths for regular words to themselves, and (ii) restricts language model paths for decomposed segments for non-lexical entities, training a n-gram language model over the training data, composing the restriction model and the language model to obtain a restricted language model, and constructing a decoding network by composing a context dependency model and a pronunciation lexicon with the restricted language model.