Local Speech-to-Text Model Bandwidth Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current voice control systems require significant bandwidth to transmit voice commands to remote servers for transcription, which is inefficient and may not function well in low-bandwidth environments.
Innovation Solution
Implementing a local speech-to-text model on devices that allows for the storage and processing of voice queries locally, reducing the need for extensive data transmission by associating audio clips with transcriptions and using these to recognize subsequent queries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If voice commands are transmitted to a remote server for transcription, then recognition accuracy is improved, but network bandwidth consumption increases
Solution Approach 1:
The patent segments the speech-to-text processing function into two parts: a lightweight local model for initial transcription and a remote server model for verification and complex queries. This segmentation allows the system to maintain high recognition accuracy while reducing bandwidth consumption by processing only necessary queries remotely.
Solution Approach 2:
The system performs preliminary transcription using the local speech-to-text model before transmitting to the remote server. This preliminary action filters out simple queries that can be handled locally, reducing the amount of data that needs to be transmitted over the network while maintaining overall recognition accuracy.
2Speed
If a comprehensive speech-to-text model is stored locally, then voice query processing speed is improved, but device memory requirements increase
Solution Approach 1:
The patent applies local quality by implementing a lightweight speech-to-text model specifically optimized for mobile devices with limited memory. This localized version of the model maintains sufficient transcription capability for common queries while occupying minimal device storage space.
Solution Approach 2:
The system uses a simplified copy of the remote server's speech-to-text model that is optimized for local deployment. This copied model retains essential transcription functionality while being compressed to fit within device memory constraints, enabling fast local processing without requiring the full server-grade model.
3Adaptability or versatility
If the local speech-to-text model is continuously updated with new transcriptions, then recognition versatility is improved, but device storage requirements increase
Solution Approach 1:
The system implements a selective update mechanism that discards redundant transcription data and retains only essential new vocabulary and patterns. When the local model is updated with new transcriptions from the remote server, it selectively incorporates only the most valuable information, maintaining versatility while controlling storage growth.
Solution Approach 2:
Instead of continuously updating the local model with all available transcriptions, the system performs partial updates only when necessary and only with the most critical new information. This partial action approach maintains recognition versatility while preventing excessive storage consumption.
Data Source
AI summary
Systems and methods are described herein for enabling, on a local device, a voice control system that limits the amount of data needed to be transmitted to a remote server. A data structure is built at the local device to support a local speech-to-text model by receiving a query and transmitting, to a remote server over a communication network, a request for a speech-to-text transcription of the query. The transcription is received from the remote server and stored in the data structure at the local device in association with an audio clip of the query. Metadata describing the query is used to train the local speech-to-text model to recognize future instances of the query.