Acoustic Echo Cancellation Using Pre-Call Training Profiles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Acoustic echo cancellers face challenges in accurately estimating and removing acoustic echoes in full duplex communication systems due to non-uniform frequency distributions and non-stationary speech signals, leading to divergence issues that require complex double filtering and may not reliably select the best estimation.
Innovation Solution
Pre-call training sessions using machine-generated audio signals to establish a trained profile, which is used to generate and adapt an acoustic echo cancellation profile during communication sessions, with a divergence detector resetting the profile if divergence exceeds a threshold, thereby reducing computational complexity and improving echo estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pre-call training sessions are used to establish a trained profile, then echo estimation accuracy is improved, but device complexity increases due to additional training system components
Solution Approach 1:
The patent implements pre-call training sessions that execute before actual communication sessions to establish a trained profile of the acoustic environment. This preliminary action captures impulse responses and frequency responses in advance, allowing the adaptive filter to start with pre-computed echo estimates rather than calculating them in real-time during calls, thereby improving accuracy without significantly increasing operational complexity
Solution Approach 2:
The training system creates a copied representation of the acoustic environment (trained profile) that can be reused across multiple communication sessions. Instead of continuously adapting during each call, the system uses this copied profile as a foundation, reducing computational load while maintaining accurate echo cancellation through the adaptive filter
2Productivity
If adaptive filtering is used during communication sessions, then computational efficiency is improved, but reliability decreases due to potential divergence from the trained profile
Solution Approach 1:
The patent incorporates a divergence detector that continuously monitors the adapted profile during communication sessions and compares it against the original trained profile. When divergence exceeds a predetermined threshold, the system provides feedback by resetting the adapted profile back to the trained profile, preventing unreliable echo cancellation while maintaining computational efficiency through adaptive filtering
Solution Approach 2:
The system dynamically adjusts between using the trained profile and the adapted profile based on divergence conditions. The adaptive filter operates dynamically during communication sessions, continuously adapting to changing acoustic conditions while the divergence detector provides dynamic monitoring and reset capability when reliability thresholds are breached
3Reliability
If double filtering methods are implemented to handle divergence, then reliability is improved, but device complexity and computational load increase significantly
Solution Approach 1:
Instead of implementing full double filtering which requires maintaining and processing multiple complete echo cancellation profiles, the patent applies partial action by using a single trained profile as the foundation and allowing limited adaptive adjustments during calls. The divergence detector provides selective intervention only when necessary, avoiding the excessive computational complexity of continuous dual-profile management while maintaining sufficient reliability
Data Source
AI summary
Disclosed methods and systems measure acoustic responses to training signals activated prior to communication sessions. Profiles related to the acoustic responses are saved and adapted during communication sessions. Training signals may have uniform frequency distributions over a frequency range and may be in response to user inputs, timeouts, or predetermined events. In the next excessive divergence is detected, an adapted profile may be substituted by an original, trained profile.