Dynamic Constrained Speech Model Selection for Aviation ASR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Automatic Speech Recognition (ASR) systems in aviation face challenges with high memory usage and latency due to the need for large vocabularies of unique keywords and acronyms, which are detrimental to recognition accuracy and latency, especially in resource-constrained cockpit environments with limited external cloud support.
Innovation Solution
A transcription system that dynamically selects and generates constrained speech models based on flight events and phase transitions, using an event detector and intelligent keyword generator to create a wordlist and switch to pre-built or newly generated models that cover specific keywords, optimizing memory usage and improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large vocabularies of unique keywords and acronyms are stored in the ASR model to improve transcription accuracy, then recognition accuracy is improved, but memory usage increases and latency increases
Solution Approach 1:
The patent segments the large aviation vocabulary into multiple constrained vocabularies based on flight phases (e.g., takeoff, cruise, landing). Each constrained vocabulary contains only the keywords relevant to a specific flight phase, reducing the memory footprint of the ASR model while maintaining transcription accuracy for the current phase.
Solution Approach 2:
The patent implements dynamic switching between different constrained vocabularies based on the detected flight phase. The system adapts the vocabulary size and content in real-time, using small vocabularies during most phases and expanding only when necessary, thereby optimizing the balance between memory usage and transcription accuracy.
2Measurement precision
If large vocabularies of unique keywords and acronyms are stored in the ASR model to improve transcription accuracy, then recognition accuracy is improved, but latency increases
Solution Approach 1:
The patent segments the large aviation vocabulary into multiple constrained vocabularies based on flight phases. Each constrained vocabulary contains only the keywords relevant to a specific flight phase, reducing the search space and processing time during ASR inference, thereby reducing latency while maintaining accuracy for the current phase.
Solution Approach 2:
The patent uses only the necessary portion of the vocabulary at any given time (partial action) rather than loading the entire vocabulary. By activating only the constrained vocabulary relevant to the current flight phase, the system reduces processing overhead and latency without sacrificing transcription accuracy for that phase.
3Quantity of substance
If constrained speech models are used to reduce memory usage and latency, then resource efficiency is improved, but the system must dynamically switch between models based on flight events
Solution Approach 1:
The patent employs a flight phase detector that continuously monitors flight parameters and provides feedback to the ASR system. Based on this feedback, the system automatically switches between constrained vocabularies, eliminating the need for manual intervention and simplifying the overall system architecture despite the dynamic model switching.
Solution Approach 2:
The ASR system automatically selects and switches between constrained vocabularies based on flight phase detection, without requiring external control or manual configuration. This self-service mechanism reduces the operational complexity of managing multiple models while maintaining resource efficiency.
Data Source
AI summary
Systems and methods are provided for the selection of a speech model for automatic speech recognition during the runtime of a transcription system, the system includes an event detector to determine one of a number of flight events that include flight plan changes and phase transitions based on data received from a set of inputs; an intelligent keyword generator to collate a set of keywords associated with the flight plan information and to generate a wordlist in response to a determination by the event detector of flight plan changes or flight phase transitions; and a processor to determine whether the wordlist is covered by a current speech model implemented in the automatic speech recognition wherein if the wordlist is not covered by the current speech model, then the processor to select a pre-built speech model that covers the wordlist for use as the current speech model in the automatic speech recognition.