Acoustic Echo Cancellation via Far-End Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing acoustic feedback control systems face challenges in effectively reducing echo and feedback in public address systems, particularly due to the difficulty in learning the echo path model during feedback situations, which often result in unstable and slow learning processes.
Innovation Solution
The system employs a method where a processor receives reference and presenter audio signals to form a trained acoustic model, allowing for concurrent echo and feedback cancellation by processing near-end audio signals, even before a participant is selected to speak, thereby creating a learning environment that optimizes echo path learning during the far-talking state and adapts to feedback situations by switching between echo and feedback cancellation modes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional acoustic feedback control systems are used to reduce echo and feedback, then echo and feedback reduction is achieved, but the learning process becomes unstable and slow
Solution Approach 1:
The system performs preliminary actions by training acoustic models during far-talking states (when only the far-end speaker is active) before actual feedback cancellation is needed. This preliminary training establishes stable echo path models in advance, so that when feedback cancellation operates during near-end talking, the models are already optimized and stable, avoiding the instability and slowness of real-time learning during feedback situations
2Adaptability or versatility
If the system processes audio signals for multiple participants, then coverage of participants is improved, but computational complexity increases
Solution Approach 1:
The system segments the acoustic environment into multiple independent acoustic models, one for each participant location. Each acoustic model processes audio signals independently for its corresponding participant, allowing the system to handle multiple participants simultaneously. This segmentation divides the complex multi-participant problem into manageable independent sub-problems, maintaining computational feasibility while improving adaptability to cover multiple participants
Data Source
AI summary
A public address system includes audio inputs from a moderator/presenter and from one or more participants. In an embodiment, the moderator speaks first and then selects participants to speak, utilizing audio captured by participant devices. A central signal processor is configured to receive the audio inputs and to utilize a configured acoustic model to provide for acoustic echo cancellation (AEC) and feedback control (FBC) during various phases of a presentation or conference. Audio signals from the presenter and/or participants, that have been processed to remove echo, are utilized as reference signals during various phases of the audio presentation that utilize the acoustic model for either AEC or FBC. The system utilizes the knowledge that the best learning occurs during the far talking state to learn the echo path in the canceler mode (AEC) vs. the feedback mode (FBC), which usually can only train in a double-talking mode.