Bidirectional Language Model for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face challenges in accurately recognizing continuous speech due to data sparseness and unidirectional language modeling, which leads to incorrect sentence composition and recognition errors, especially with non-grammatically correct utterances and distant word correlations.
Innovation Solution
A speech recognition apparatus and method that utilizes bidirectional language models to calculate scores for word strings, selecting between forward and backward language models to account for bidirectional relationships between words, and integrates acoustic model scores to output high-scoring sentences, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional unidirectional language models are used to reduce computation, then processing speed improves, but recognition accuracy deteriorates for non-grammatical utterances and distant word correlations
Solution Approach 1:
The language model is segmented into forward language model and backward language model components. The forward language model processes words in forward direction (P(word_i|word_{i-1}, ..., word_{i-n+1})), while the backward language model processes words in backward direction (P(word_i|word_{i+1}, ..., word_{i+n-1})). This segmentation allows the system to capture bidirectional contextual relationships without requiring computation for all possible word combinations, thus improving recognition accuracy while maintaining processing efficiency.
Solution Approach 2:
The patent introduces a bidirectional dimension to the traditional unidirectional language modeling. By adding the backward language model dimension, the system considers contextual relationships from both forward and backward directions, enabling accurate recognition of non-grammatical utterances and distant word correlations that unidirectional models miss.
2Measurement precision
If all word combinations are modeled to improve accuracy, then recognition precision improves, but computation complexity increases exponentially
Solution Approach 1:
Instead of modeling all possible word combinations (excessive action), the patent applies partial action by only modeling forward and backward directional relationships. This selective approach captures the essential contextual dependencies needed for accurate speech recognition while avoiding the exponential computation complexity of complete word combination modeling.
Solution Approach 2:
The comprehensive language modeling task is segmented into manageable forward and backward components. Each component models specific directional relationships, dividing the complex computation into simpler sub-tasks that can be processed efficiently without requiring exponential computational resources.
3Productivity
If language models only consider adjacent words to simplify processing, then computational efficiency improves, but recognition accuracy deteriorates for distant word correlations
Solution Approach 1:
The patent adds a backward directional dimension to complement the forward direction. This bidirectional approach enables the model to capture distant word correlations by considering contextual relationships from both directions, improving recognition accuracy while maintaining computational efficiency through the structured bidirectional framework.
Data Source
AI summary
A speech recognition apparatus is provided. The speech recognition apparatus includes a primary speech recognition unit configured to perform speech recognition on input speech and thus to generate word lattice information, a word string generation unit configured to generate one or more word strings based on the word lattice information, a language model score calculation unit configured to calculate bidirectional language model scores of the generated word strings selectively using forward and backward language models for each of words in each of the generated word strings, and a sentence output unit configured to output one or more of the generated word strings with high scores as results of the speech recognition of the input speech based on the calculated bidirectional language model scores.