Voice Recognition Graph Partitioning for Memory-Constrained Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing navigation and location-based systems face challenges in implementing automatic speech recognition due to resource constraints, particularly in embedded and mobile devices, where the large global speech decoding graph exceeds memory capacity, leading to out-of-vocabulary errors and poor user experience.
Innovation Solution
The system partitions the global speech decoding graph into spatial partitions, determines key entities within each partition, constructs a combined set of key entities, and creates a retrieval index to map these entities to corresponding partitions, allowing for seamless re-processing of voice inputs when out-of-vocabulary results occur, thereby enhancing natural voice-based entry of geographic information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a global speech decoding graph is used to provide comprehensive voice recognition, then voice recognition accuracy is improved, but device memory requirements exceed available capacity
Solution Approach 1:
The global speech decoding graph is divided into multiple regional partitions, each stored separately in device memory. This segmentation allows the system to manage memory resources efficiently by loading only relevant regional graphs rather than attempting to store the entire global graph at once.
Solution Approach 2:
The system pre-loads speech decoding graphs for regions that are geographically close to the device's current location into memory before they are needed for voice recognition. This preliminary action ensures that when users speak destination names in their current or nearby regions, the appropriate decoding graphs are already available, maintaining recognition accuracy without requiring all global data to be simultaneously stored.
2Quantity of substance
If regional speech decoding graphs are stored to save memory, then device memory usage is reduced, but out-of-vocabulary errors increase for destinations outside stored regions
Solution Approach 1:
When the speech recognition system encounters an out-of-vocabulary error (indicating the spoken destination is not in the currently loaded regional graph), it provides feedback by identifying the unrecognized location and automatically loading the corresponding regional speech decoding graph into memory. This feedback mechanism ensures that recognition reliability is maintained dynamically without requiring all possible regions to be pre-loaded.
Solution Approach 2:
The system dynamically adjusts the set of loaded speech decoding graphs based on the device's current location and the user's voice input needs. Rather than using a static set of pre-loaded regions, the system adapts by loading additional regional graphs as needed, optimizing the balance between memory usage and recognition reliability in real-time.
3Adaptability or versatility
If the speech decoding graph is updated to include new locations, then voice interaction versatility is improved, but the graph size increases beyond device capacity
Solution Approach 1:
The speech decoding graph is organized into modular regional segments that can be independently loaded and unloaded. When new locations are added to the global graph, they are incorporated as separate regional partitions rather than expanding a monolithic structure. This allows the system to support global voice interaction versatility while maintaining manageable graph sizes through selective loading of only the regions currently needed.
Data Source
AI summary
An approach is provided for global voice-based entry of location information. The approach involves partitioning a global speech decoding graph into spatial partitions. The approach also involves determining key entities occurring in each spatial partitions to construct a combined set of key entities. The approach further involves creating a retrieval index to map the key entities in the combined set of key entities to a corresponding partition. A first partition, the combined set of key entities, and the retrieved index are stored in a memory of a device for processing a voice input signal. A second partition that is not in the memory of the device is retrieved based on the combined set of key entities and the retrieval index to automatically re-process the voice input signal when an out-of-vocabulary result is obtained from the first partition.