WFST Random Access Compression for On-Board Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems on small devices, such as wearables and smartphones, face memory constraints due to the computational complexity and large memory requirements of weighted finite state transducers (WFSTs), limiting vocabulary size and requiring remote computations, which result in delays and increased battery usage.
Innovation Solution
A random access compression technique for WFSTs that stores arcs in separate integer arrays, allowing flexible compression and decompression of individual elements, reducing memory requirements by approximately 30% for small vocabularies and 38% for large vocabularies, enabling on-board speech recognition without the need for remote computations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If WFSTs are stored in uncompressed format, then speech recognition accuracy is maintained, but memory requirements exceed available capacity on small devices
Solution Approach 1:
The patent segments the WFST data structure into separate integer arrays (e.g., source states, destination states, input labels, output labels, weights) that can be independently compressed and stored. This segmentation allows for efficient compression algorithms to be applied to each array individually, reducing overall memory requirements while preserving the complete transducer information needed for accurate speech recognition.
Solution Approach 2:
The patent transforms the WFST parameters from their original uncompressed integer representations into compressed formats using techniques such as differential encoding, where only the differences between consecutive values are stored. This parameter transformation significantly reduces the storage space required while maintaining the ability to reconstruct the original values for accurate speech recognition decoding.
2Adaptability or versatility
If remote server computations are used, then large vocabulary speech recognition is achieved, but communication delays and battery consumption increase
Solution Approach 1:
The patent extracts the computationally intensive WFST data structure from remote server environments and brings it locally to small devices through efficient compression. By taking out the previously server-dependent vocabulary and grammar representations and storing them in compressed form on-device, the system enables offline speech recognition with large vocabularies, eliminating communication delays and reducing battery consumption from wireless transmissions.
3Quantity of substance
If WFSTs are compressed to reduce memory usage, then on-board processing becomes feasible, but decompression overhead may increase processing time
Solution Approach 1:
The patent applies preliminary compression to the WFST data structure during system initialization or offline setup, transforming the data into a compact integer array format. This preliminary action ensures that when speech recognition needs to occur, the compressed data is already prepared and can be efficiently decompressed on-demand, minimizing the impact on real-time processing speed while achieving significant memory savings.
Solution Approach 2:
The patent replaces traditional mechanical decompression approaches with optimized integer array manipulations and mathematical transformations. By substituting general-purpose decompression mechanisms with specialized integer-based operations that leverage the structured nature of WFST data, the system achieves faster decompression rates that maintain real-time processing capabilities despite the compression overhead.
Data Source
AI summary
A system, article, and method of random access compression of transducer data for automatic speech recognition decoding.