Hierarchical Networks for Efficient Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) engines face challenges in achieving both efficiency and flexibility, with prior art systems either being inefficient due to redundant computations or inflexible due to static designs, limiting their ability to adapt to different tasks and languages.
Innovation Solution
The use of hierarchical networks that can be dynamically created and modified during recognition, incorporating a result network, network tokens, and a network model, allows for efficient computations by consolidating redundant processes and flexibility by adapting to different applications and languages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LVCSR systems increase in size to handle larger vocabularies and improve accuracy, then speech recognition accuracy is improved, but computing resources and time required for the search increase significantly
Solution Approach 1:
The ASR engine is divided into multiple parallel search components (first search component, second search component, etc.) that operate simultaneously on different portions of the hypothesis space. Each component maintains its own hypothesis list and performs independent search operations, allowing the system to process large vocabularies without requiring a single monolithic search structure that would consume excessive resources.
Solution Approach 2:
The invention introduces a new dimensional organization to the search space by creating multiple parallel search components instead of expanding a single search structure. This dimensional change allows the system to handle larger vocabularies by distributing the search load across multiple independent threads, each managing a subset of hypotheses, thereby improving computational efficiency while maintaining accuracy.
2Measurement precision
If the ASR engine performs a comprehensive search among all possible utterances to improve accuracy, then speech recognition accuracy is improved, but the time required to perform speech recognition increases
Solution Approach 1:
The search space is segmented into multiple parallel search components, each handling a portion of the hypothesis space. By distributing the search workload across multiple independent components that operate simultaneously, the system can explore more hypotheses comprehensively without linearly increasing the time required, as parallel execution reduces the total recognition time compared to sequential processing.
Solution Approach 2:
Multiple search components continuously perform search operations in parallel without idle periods. Each component maintains an active hypothesis list and continuously evaluates acoustic and language model scores, ensuring that useful computational work is performed at all times across all parallel threads, maximizing resource utilization and reducing overall recognition time.
3Productivity
If early ASR systems used simple models to reduce computing resources, then computing efficiency is improved, but vocabulary size and recognition capability are limited
Solution Approach 1:
The system segments the large vocabulary search space into multiple manageable portions handled by separate search components. Each component works with a subset of hypotheses and maintains its own data structures, allowing the system to support large vocabularies without requiring a single complex data structure that would consume excessive memory and processing resources.
Solution Approach 2:
The parallel search component architecture provides a universal framework that can handle various vocabulary sizes and language complexities. Each search component is designed to be self-contained and can be configured for different language models and acoustic models, making the system adaptable to multiple languages and vocabulary requirements while maintaining computing efficiency through standardized parallel processing.
Data Source
AI summary
Provided are systems and methods for using hierarchical networks for recognition, such as speech recognition. Conventional automatic recognition systems may not be both efficient and flexible. Recognition systems are disclosed that may achieve efficiency and flexibility by employing hierarchical networks, prefix consolidation of networks, and future consolidation of networks. The disclosed networks may be associated with a network model and the associated network model may be modified during recognition to achieve greater flexibility.