Source Speech Recognition with Acoustic Model Negotiation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face high word error rates, especially in free-form speech recognition, due to the need for large vocabularies and computational resources, and struggle with user-specific speech variations such as accents and dialects, leading to inefficient performance in mobile devices and centralized systems.
Innovation Solution
A dynamic negotiation between user devices and destination servers for speech recognition services, where devices can provide speech attributes and acoustic models, allowing for partial or full recognition on the device or server, using protocols like RTP and MLRR algorithms to adapt and improve recognition accuracy over time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If speech recognition is performed on user devices with limited computational resources, then device autonomy and response time are improved, but recognition accuracy deteriorates due to insufficient processing power
Solution Approach 1:
The speech recognition system is segmented into multiple components: acoustic model processing, language model processing, and decoding. The acoustic model is downloaded to the mobile device and processed locally, while language models and complex decoding operations can be distributed to remote servers. This segmentation allows computationally intensive tasks to be performed on devices with appropriate resources while maintaining local processing for time-critical functions.
Solution Approach 2:
The patent introduces an intermediary architecture where a remote server acts as a mediator between the mobile device and the speech recognition processing. The server provides supplementary language models, vocabulary extensions, and complex acoustic processing that the mobile device cannot handle alone, while the device handles initial processing and time-critical recognition tasks locally.
2Measurement precision
If large vocabulary and computational resources are allocated to speech recognition systems, then free-form speech recognition accuracy is improved, but device complexity and resource requirements increase
Solution Approach 1:
The system dynamically adapts its resource allocation and processing architecture based on available resources and task requirements. Acoustic models are downloaded on-demand, language model complexity is adjusted based on vocabulary size requirements, and processing is dynamically distributed between local and remote components. This dynamic approach allows the system to scale accuracy with resources when available while maintaining functionality with limited resources.
Solution Approach 2:
The patent changes key parameters of the speech recognition system including vocabulary size, language model order, and acoustic model complexity based on available computational resources. The system can operate with reduced vocabulary and simpler models on mobile devices while achieving higher accuracy with expanded parameters when connected to remote servers or running on more powerful platforms.
3Power
If centralized systems are used for speech recognition, then computational power and training capabilities are improved, but adaptability to user-specific speech patterns deteriorates due to limited training opportunities
Solution Approach 1:
The system implements feedback mechanisms where speech recognition results and user corrections are continuously fed back to update acoustic models and language models. User-specific speech patterns are captured through repeated interactions, and the system adapts by updating personalized models that are downloaded to the mobile device. This feedback loop enables centralized computational power to be leveraged while achieving user-specific adaptation through iterative learning.
Solution Approach 2:
The system performs preliminary actions by pre-downloading acoustic models and language models to the mobile device before speech recognition tasks are executed. User-specific adaptations are pre-computed and stored as personalized models that are downloaded in advance, enabling the device to operate with enhanced accuracy without requiring real-time connection to centralized systems.
Data Source
AI summary
Recognizing a user's speech is a computationally demanding task. If a user calls a destination server, little may be known about the user or the user's speech profile. The user's source system (device and/or server) may have an extensive profile of the user. As provided herein, a source device may provide translated text and/or speech attributes to a destination server. As a benefit, the recognition algorithm may be well tuned to the user and provide the recognized content to the destination. Additionally, the destination may provide domain attributes to allow the source recognition engine to better recognize the spoken content.