Acoustic Language Model Training via Word Class Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face limitations in vocabulary size due to hardware and software constraints, leading to decreased recognition accuracy for low-frequency words and out-of-glossary terms, as the available vocabulary size is capped at around one hundred thousand, restricting the expansion of vocabulary coverage.
Innovation Solution
A method and device for training an acoustic language model that uses word segmentation and class replacement to reduce the total vocabulary count by replacing entity names with word class labels, allowing for improved recognition accuracy by expanding vocabulary coverage without increasing the model size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the vocabulary size of the language model is increased to cover more words, then the vocabulary coverage is improved, but the model size grows exponentially and exceeds hardware limitations
Solution Approach 1:
The patent merges multiple specific entity names that belong to the same word class into a single generic word class label. For example, specific names like 'Beijing', 'Shanghai', 'Guangzhou' are merged into the location word class label 'LOC'. This consolidation reduces the vocabulary size while maintaining the ability to represent diverse entities, thereby resolving the contradiction between vocabulary coverage and model size.
Solution Approach 2:
The patent creates universal word class labels that can represent multiple specific entities. A single word class label (e.g., 'ORG' for organization, 'LOC' for location) serves multiple functions by standing in for numerous specific entity names. This multi-functionality allows the model to handle a wide vocabulary range without proportionally increasing model size, as the same word class label can substitute for many different specific words.
2Reliability
If the language model includes all possible words to improve recognition accuracy, then the vocabulary coverage increases, but the computational complexity and processing time increase significantly
Solution Approach 1:
By merging specific entity names into generic word class labels, the patent reduces the number of unique vocabulary items the model must process. Instead of computing probabilities for thousands of specific entity names, the model computes probabilities for a smaller set of word class labels, thereby reducing computational complexity while maintaining recognition accuracy through the representative power of word class labels.
Solution Approach 2:
Universal word class labels serve as compact representations that reduce computational burden. When the model encounters any specific entity name, it can use the corresponding word class label for processing, avoiding the need to handle each specific name individually. This universality simplifies the computational process while preserving the ability to recognize diverse entities accurately.
3Adaptability or versatility
If the model is trained with a large vocabulary to cover low-frequency words, then the vocabulary coverage improves, but the training data requirements and training time increase exponentially
Solution Approach 1:
The patent merges training data for multiple specific entity names under a single word class label. Instead of requiring separate training examples for each specific entity name (which would exponentially increase training data requirements), the model consolidates training data at the word class level. This approach reduces training time while still enabling the model to recognize specific entities through their word class labels.
Solution Approach 2:
Word class labels provide a universal training target that aggregates statistics from multiple specific entities. During training, the model learns the probability distributions of word class labels from training corpora, and this learned knowledge can be applied to recognize any specific entity belonging to that class. This universality eliminates the need for extensive training on each specific entity, dramatically reducing training time while maintaining vocabulary coverage.
Data Source
AI summary
A method and a device for training an acoustic language model, include: conducting word segmentation for training samples in a training corpus using an initial language model containing no word class labels, to obtain initial word segmentation data containing no word class labels; performing word class replacement for the initial word segmentation data containing no word class labels, to obtain first word segmentation data containing word class labels; using the first word segmentation data containing word class labels to train a first language model containing word class labels; using the first language model containing word class labels to conduct word segmentation for the training samples in the training corpus, to obtain second word segmentation data containing word class labels; and in accordance with the second word segmentation data meeting one or more predetermined criteria, using the second word segmentation data containing word class labels to train the acoustic language model.