Accumulative Multi-Cue Activation for Domain-Specific ASR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing virtual assistant applications in mobile devices require a wake-up word to activate the full automatic speech recognition (ASR) engine, leading to increased power consumption and shorter battery life due to the engine's high memory footprint and processing cycles.
Innovation Solution
The implementation of an accumulative multi-cue activation (AMCA) module that selectively activates a domain-specific ASR engine based on detected keywords and context words, allowing for speech-to-text conversion without a wake-up word, thereby reducing computational and power resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a full ASR engine is constantly activated to detect general speech in natural language, then speech recognition capability is improved, but power consumption increases and battery life decreases
Solution Approach 1:
The ASR system is segmented into multiple domain-specific engines (e.g., weather domain, sports domain, news domain) instead of using a single full ASR engine. Each domain-specific engine is optimized for particular keywords and contexts, allowing the system to activate only the relevant engine for each query, thereby reducing overall power consumption while maintaining recognition capability.
Solution Approach 2:
Instead of activating the complete ASR engine for all queries, the system activates only the necessary portion (domain-specific engine) based on keyword matching and domain determination. This partial action approach activates just enough processing power to handle the specific query type, avoiding unnecessary computation and energy waste.
2Measurement precision
If a full ASR engine is activated to process speech input without a wake-up word, then response accuracy is improved, but computational resources and memory footprint increase
Solution Approach 1:
The system segments the ASR processing into domain-specific modules, each optimized for particular types of queries. This segmentation allows the system to use simpler, more efficient processing for domain-specific queries while maintaining high accuracy for those domains, rather than always deploying the computationally intensive full ASR engine.
Solution Approach 2:
An intermediary component (domain determination module) is introduced between the keyword detection and ASR activation. This intermediary analyzes the query context and determines the appropriate domain, then activates only the corresponding domain-specific ASR engine. This intermediary layer enables precise activation decisions that balance accuracy requirements with computational efficiency.
3Use of energy by moving object
If a wake-up word is required to activate the ASR engine, then power consumption is reduced, but user convenience decreases when the wake-up word is forgotten
Solution Approach 1:
The system performs preliminary keyword detection and domain determination before activating the ASR engine. By pre-processing the query to identify domain-specific keywords and contexts, the system can make informed decisions about engine activation without requiring a wake-up word, thus maintaining both energy efficiency and user convenience.
Solution Approach 2:
The system uses the content of the user's query itself (keywords and context) to trigger the appropriate ASR engine, rather than requiring a separate wake-up word. The query's own linguistic features serve as the activation mechanism, allowing the system to be both energy-efficient and convenient by eliminating the need for an additional triggering phrase.
Data Source
AI summary
A method includes detecting a keyword within an audio stream. The keyword is one of multiple keywords in a database, in which each of the multiple keywords relates to at least one of multiple domains in the database. The database stores a first confidence weight for each of the multiple keywords that are related to a first domain among the multiple domains. Each first confidence weight indicates a probability that a corresponding keyword relates to the first domain. The method includes determining whether a first confidence weight of the keyword is at least equal to an activation threshold value associated with the first domain. The method includes, in response to the first confidence weight of the keyword meeting the activation threshold value, activating a DS-ASR engine corresponding with the first domain to perform speech-to-text conversion on the audio stream.