Decoder Token Duplication Elimination for WFST Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing decoder technologies face challenges in efficiently handling self-transitions in hidden Markov models (HMMs) when integrated with weighted finite state transducers (WFSTs), leading to increased memory requirements and complexity due to the need for modifying WFSTs to handle multiple input symbols for each state, which complicates speech recognition processes.
Innovation Solution
A decoder is designed with a token operating unit and a duplication eliminator that propagates tokens and eliminates duplicate tokens with the same state and input symbol, allowing for efficient search in WFSTs without expanding the WFST to have single-type input symbols for each state, thereby reducing the number of states and transitions and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the WFST is modified to handle self-transitions by limiting transitions to single input symbol types per state, then the HMM can be correctly dealt with, but the number of states and transitions increases, leading to larger memory requirements
Solution Approach 1:
The patent segments the token management process into two distinct components: a token operating unit that propagates tokens through the WFST, and a duplication eliminator that removes duplicate tokens. This segmentation allows the system to handle multiple input symbol types at each state without requiring additional states or transitions, thereby maintaining a compact WFST structure while correctly processing HMM self-transitions.
Solution Approach 2:
The duplication eliminator acts as an intermediary between the token operating unit and the final output. It receives tokens propagated by the token operating unit, eliminates duplicates based on state and input symbol matching, and passes the refined token set forward. This intermediary component enables correct HMM handling without expanding the WFST structure.
2Reliability
If the WFST is expanded to have single-type input symbols for each state to handle self-transitions, then HMM processing is correct, but memory capacity requirements increase
Solution Approach 1:
By dividing the processing function into token propagation and duplication elimination stages, the system avoids the need to expand the WFST structure. The original compact WFST is preserved, maintaining low memory requirements, while the segmented processing logic ensures accurate HMM self-transition handling.
Solution Approach 2:
The system creates virtual copies of tokens with different input symbol types at the same state without creating additional physical states or transitions in the WFST. The duplication eliminator then manages these virtual copies, allowing accurate HMM processing while keeping the actual WFST structure compact and memory-efficient.
3Ease of operation
If separate data structures are maintained for token assignments and HMMs, then detailed control is possible, but device complexity and management overhead increase
Solution Approach 1:
The patent merges the token assignment and HMM processing data structures into a unified token structure. Each token contains both the state information and the input symbol type, allowing the system to maintain detailed control over the recognition process while using a single integrated data structure. This eliminates the need to manage separate data structures for token assignments and HMMs, reducing complexity while preserving control precision.
Data Source
AI summary
According to an embodiment, a decoder searches a finite state transducer and outputs an output symbol string corresponding to a signal that is input or corresponding to a feature sequence of signal that is input. The decoder includes a token operating unit and a duplication eliminator. The token operating unit is configured to, every time the signal or the feature is input, propagate each of a plurality of tokens, which is assigned with a state of the head of a path being searched, according to the finite state transducer. The duplication eliminator is configured to eliminate duplication of two or more tokens which have same state assigned thereto and for which respective previously-passed transitions are assigned with same input symbol.