Hybrid Speech Recognition Segmentation for Latency and Vocabulary
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems for embedded devices, such as mobile phones and automotive systems, face limitations in handling large vocabularies, multiple languages, and real-time processing under noisy conditions due to insufficient processing power and bandwidth constraints, as well as reliance on fixed dictionaries and language models.
Innovation Solution
A method that segments speech utterances and assigns processing tasks between a thin client device and remote servers, utilizing multiple speech recognizers with adaptable language and acoustic models, allowing for distributed processing and improved accuracy and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If speech recognition is performed entirely on embedded devices, then real-time processing is achieved, but vocabulary size and language support are limited
Solution Approach 1:
The speech recognition task is segmented into multiple parts: initial analysis and segmentation of speech utterance is performed locally on the embedded device, while more computationally intensive processing (acoustic modeling, language model matching) is distributed to remote servers. This allows the system to achieve both real-time responsiveness for initial processing and comprehensive vocabulary/language support through server-based resources.
2Measurement precision
If multiple speech recognizers are deployed on embedded devices, then recognition accuracy improves, but processing power requirements increase
Solution Approach 1:
A hybrid client-server architecture is introduced as an intermediary between the embedded device and multiple speech recognizers. The local device performs initial speech segmentation and sends relevant segments to remote servers that host multiple speech recognizers. This intermediary approach enables access to multiple recognizers for improved accuracy without requiring the embedded device to have the processing power to run all recognizers locally.
3Adaptability or versatility
If speech processing is centralized on servers, then vocabulary and language capabilities increase, but bandwidth consumption increases
Solution Approach 1:
Instead of sending entire speech utterances to servers for processing, the system performs partial processing locally by segmenting the speech and identifying relevant segments. Only these segmented portions are transmitted to servers for further processing. This partial action approach reduces bandwidth consumption while still enabling servers to provide comprehensive language and vocabulary capabilities.
4Adaptability or versatility
If embedded devices have sufficient processing power for comprehensive speech recognition, then recognition capabilities improve, but device complexity and cost increase
Solution Approach 1:
The system implements a universal hybrid architecture where embedded devices serve as thin clients that can access speech recognition capabilities both locally and remotely. This multi-functional design allows the same device to operate in different modes (local-only, server-assisted, or fully cloud-based) depending on available resources and requirements, avoiding the need for each device to have dedicated high-end processing hardware.
Data Source
AI summary
A recipient computing device can receive a speech utterance to be processed by speech recognition and segment the speech utterance into two or more speech utterance segments, each of which can be to one of a plurality of available speech recognizers. A first one of the plurality of available speech recognizers can be implemented on a separate computing device accessible via a data network. A first segment can be processed by the first recognizer and the results of the processing returned to the recipient computing device, and a second segment can be processed by a second recognizer implemented at the recipient computing device.