On-the-fly Word Lattice Generation for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face inefficiencies in generating word lattices, which require significant RAM and slow down the recognition process due to the need for large state or phone lattices, necessitating a more efficient approach.
Innovation Solution
The system generates word lattices in a single pass during decoding without building intermediate state or phone lattices, using a weighted finite-state transducer (WFST) to create a word lattice independently of WFST decoding, and limits token recombination based on word history designations to reduce memory usage and maintain decoding speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If word lattices are built from state or phone lattices in a second step, then word lattice generation is achieved, but RAM requirements increase significantly and processing time increases
Solution Approach 1:
The patent merges the word lattice generation process with the WFST decoding process by propagating tokens through the WFST and simultaneously building the word lattice in a single pass. This eliminates the need for separate state/phone lattice generation and subsequent word lattice construction, thereby reducing RAM requirements and processing time while maintaining recognition accuracy.
Solution Approach 2:
The patent performs preliminary action by generating word history designations and building the word lattice structure during the decoding process itself, rather than waiting for decoding to complete and then constructing the lattice. This proactive approach ensures the lattice is available immediately for NLU processing without additional memory or time overhead.
2Productivity
If state or phone lattices are generated on the fly during speech decoding, then speech decoding is performed, but the lattices become large and require significant RAM
Solution Approach 1:
The patent extracts only the necessary information for word lattice generation during WFST decoding, specifically propagating tokens with word history designations and building a compressed word lattice structure. This extraction approach maintains decoding speed while significantly reducing the memory footprint compared to storing complete state or phone lattices.
Solution Approach 2:
The patent creates a simplified copy of the lattice structure focused on word-level hypotheses rather than preserving the complete state or phone lattice. This copying approach retains the essential information needed for NLU processing while discarding redundant details, thereby reducing RAM requirements.
3Reliability
If word lattices are generated in a second step from state/phone lattices, then word lattice is produced, but the process becomes slower and less efficient
Solution Approach 1:
The patent ensures continuity of useful action by performing word lattice generation continuously during the WFST decoding process rather than as a separate sequential step. Tokens are propagated and lattice nodes are created simultaneously with decoding operations, eliminating idle time and ensuring lattice information is ready immediately for NLU processing.
Solution Approach 2:
The patent performs preliminary action by pre-building the word lattice structure during decoding so that it is ready for NLU processing from the start. This eliminates the time delay associated with generating lattices after decoding completes, while maintaining the accuracy and reliability of lattice information through systematic token propagation and word history tracking.
Data Source
AI summary
A systems, article, and method of automatic speech recognition using on-the-fly word lattice generation with word histories.