ATC Transcription Split Prediction for Unsplit Voice Segments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Voice Activity Detection (VAD) devices in aircraft systems fail to detect speaker changes in fast-paced ATC-pilot dialogues with pauses shorter than 20 milliseconds, leading to unsplit messages that cause transcription errors and miscommunications.
Innovation Solution
A transcription system with a VAD module and a split predictor that performs content-based analysis to identify split points in audio segments, using intelligent algorithms to determine appropriate splits based on repetitive call signs or critical information, and re-decode unsplit messages if necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional Voice Activity Detection (VAD) devices are used to monitor fast controller/pilot dialogues, then the system can handle normal speech environments, but it fails to detect speaker changes when pauses are shorter than 20 milliseconds
Solution Approach 1:
The system changes the detection parameter from fixed pause duration threshold to dynamic threshold based on dialogue context. The split predictor analyzes message content and speaker turn patterns to adaptively determine split points, allowing reliable detection even when pauses are shorter than the conventional 20ms threshold.
Solution Approach 2:
The patent introduces a split predictor as an intermediary component between the VAD device and the transcription system. This intermediary analyzes message content and uses intelligent algorithms to predict split points, bridging the gap between conventional VAD limitations and the need for accurate speaker change detection in fast-paced dialogues.
2Ease of operation
If the VAD device operates with conventional pause detection thresholds, then it functions normally in standard conversations, but it produces unsplit messages in fast-paced ATC dialogues
Solution Approach 1:
The patent segments the message processing into two stages: initial VAD processing with conventional thresholds, followed by secondary segmentation using the split predictor. This allows the system to maintain simple VAD operation while achieving precise message segmentation through content-based analysis and intelligent algorithms.
3Productivity
If the system uses conventional transcription processing, then it processes messages in real-time, but it produces transcription errors when messages are not properly segmented
Solution Approach 1:
The split predictor performs preliminary analysis of message content before final transcription processing. By predicting split points based on content analysis and intelligent algorithms, the system prepares properly segmented messages for transcription, ensuring both real-time processing speed and high transcription accuracy.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Systems and methods are provided for a transcription system with voice activity detection (VAD). The system includes a VAD module to receive incoming audio and generate an audio segment; and a speech decoder with a split predictor to perform, in a first pass, a decode operation to transcribe text from an audio segment into a message; wherein in the first pass, if the message is determined not to contain a split point based on a content-based analysis performed by the split predictor, the speech decoder forwards the message for display and if the message is determined based on the content-based analysis to contain the split point, the speech decoder performs in a second pass, a re-decode operation to transcribe text from the audio segment based on the split point wherein the split point is configured within an audio domain of the audio segment by the split predictor and forward the message for display.