Voice Recognition Model Weight Sharing for On-Device Memory Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition models face memory limitations in on-device implementations, leading to reduced processing speed and inefficiency due to limited storage space for weight and parameter data.
Innovation Solution
An electronic apparatus and method that shares weight information across different networks within a speech recognition model, optimizing memory usage by storing overlapping weight information in separate areas of the memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the speech recognition model stores all weight information in local memory, then speech recognition accuracy is maintained, but memory usage increases and processing speed decreases
Solution Approach 1:
The patent segments weight information into two categories: first weight information (unique to each network) and second weight information (shared across multiple networks). This segmentation allows the system to store only essential unique weights locally while sharing common weights, thereby reducing memory usage and improving processing speed while maintaining recognition accuracy.
Solution Approach 2:
The patent implements universality by having multiple networks share the same second weight information. The shared weight information serves multiple networks simultaneously, reducing overall memory requirements and enabling faster access during speech recognition operations without compromising the functionality or accuracy of individual networks.
2Reliability
If the speech recognition model uses more storage space for weight information, then processing accuracy is improved, but device memory limitations are exceeded
Solution Approach 1:
The patent merges common weight information across multiple networks into a single shared storage location. By combining the second weight information that is used by multiple networks into one copy rather than duplicating it across each network, the system significantly reduces total memory consumption while preserving the accuracy required for speech recognition.
Solution Approach 2:
The patent divides weight information into first weight information (stored locally for each network) and second weight information (shared across networks). This segmentation strategy ensures that only the minimum necessary unique weights are stored locally, optimizing the balance between recognition accuracy and memory storage capacity.
3Reliability
If the speech recognition model is implemented on-device, then user privacy is protected, but memory usage and storage space are limited
Solution Approach 1:
The patent enables multiple speech recognition networks to share common weight information stored on the device. This universality allows the system to run multiple networks with reduced individual storage requirements, making on-device implementation feasible within limited memory constraints while maintaining privacy by keeping all processing local.
Solution Approach 2:
The patent merges shared weight information into a single storage location that serves multiple networks. This consolidation reduces the total storage footprint required for on-device speech recognition, enabling privacy-protecting local processing without exceeding device memory limitations.
4Adaptability or versatility
If the speech recognition model keeps all parameters in local memory, then model flexibility is maintained, but processing speed becomes impractical
Solution Approach 1:
The patent segments parameters into first parameters (network-specific, stored locally) and second parameters (shared across networks, accessed efficiently). This segmentation maintains model flexibility by preserving network-specific adaptations while improving processing speed through efficient shared parameter access, avoiding the bottleneck of loading all parameters from local memory.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present electronic device comprises: a memory for storing a voice recognition model and first recognition information corresponding to a first user voice acquired through the voice recognition model, the voice recognition model including a first network, a second network and a third network; and a processor, which inputs voice data corresponding to a second user voice in the first network so as to acquire a first vector, inputs the first recognition information in the second network for generating a vector on the basis of first weight information, so as to acquire a second vector, inputs the first vector and the second vector in the third network for generating the recognition information on the basis of second weight information, so as to acquire the second recognition information corresponding to the second user voice, wherein at least a part of the second weight information is the same as the first weight information.