Spectral Domain Non-Linear Echo Cancellation in Hands-Free Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional linear adaptive echo cancellers in hands-free telephony systems fail to completely eliminate non-linear echo residuals, especially during double talk periods, due to limited DSP resources and environmental noise, leading to noticeable echo residuals in full duplex communication systems, particularly in mobile or VoIP networks.
Innovation Solution
A spectral domain non-linear echo cancellation method that enhances echo suppression by converting input signals into spectral subbands, estimating echo residual power, calculating signal-to-echo ratios, and applying gains to cancel non-linear echoes, which can be integrated with existing linear adaptive filters, reducing the need for additional non-linear processing circuits and improving full duplex operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If a linear adaptive filter is used for echo cancellation, then the echo can be suppressed, but non-linear echo residuals remain noticeable especially during double talk periods
Solution Approach 1:
The patent transforms the echo cancellation problem from time domain to spectral domain by applying FFT to convert signals into frequency components. This parameter transformation enables the system to identify and suppress non-linear echo residuals that are invisible in the time domain, achieving over 60 dB echo cancellation while preserving full duplex operation during double talk periods
Solution Approach 2:
The patent divides the spectral domain into multiple subbands using spectral subband decomposition. By segmenting the frequency spectrum into distinct bands and processing each band independently with subband-specific gains, the system can selectively suppress echo residuals in different frequency regions while maintaining speech quality, resolving the contradiction between echo suppression and full duplex reliability
2Object-affected harmful factors
If a non-linear processing (NLP) circuit is added to suppress remaining echo residual, then echo suppression improves, but the near-end talker's voice is clipped during double talk periods
Solution Approach 1:
The patent replaces the mechanical NLP circuit approach with a spectral domain processing system. Instead of using non-linear time-domain processing that clips signals during double talk, the system uses FFT-based spectral analysis with selective gain application in frequency domain, eliminating the clipping problem while achieving superior echo suppression. The spectral domain approach naturally distinguishes echo from speech without aggressive non-linear clipping
Solution Approach 2:
The patent introduces spectral subband decomposition as an intermediary between the linear adaptive filter and the final output. This intermediate spectral representation allows the system to identify echo components that survive linear cancellation and apply targeted suppression gains only to those components, avoiding the need for aggressive NLP that would clip double talk signals
3Measurement precision
If the adaptive filter size is increased to improve echo estimation accuracy, then echo cancellation performance improves, but DSP engine resources (memory and MIPS) are exceeded
Solution Approach 1:
The patent moves the problem from time domain to spectral domain using FFT transformation. This dimensional change allows the system to achieve high echo cancellation performance (over 60 dB) with a relatively small adaptive filter by exploiting frequency domain characteristics. The spectral decomposition reveals echo patterns that are not apparent in time domain, enabling accurate echo estimation with limited DSP resources
Data Source
AI summary
Disclosed is a non-linear echo canceller and method for cancelling echo during full duplex communication in a hands free communication system. An input signal from a far-end talker and an input signal from the output from an echo canceller are received. K spectral subbands are created for each input signal. The spectral echo residual power at each subband is estimated and compared to a clean signal power to calculate a signal to echo ratio. Gains are calculated based on each calculated ratio and non-linear echo is cancelled based on the calculated gains.