Spectral Double-Talk Detection for Hands-Free Echo Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing double talk detection methods in hands-free telephony systems often fail to accurately detect double talk conditions, leading to residual echo and voice clipping, especially when the echo level is higher than the near-end signal level, due to limitations in adaptive echo cancelers and reliance on signal power differences.
Innovation Solution
A method that detects double talk conditions based on inherent frequency response differences between transducers and the acoustical effect on the spectrum of the returned echo signal, using selected subbands and echo loss coefficients, to improve duplex performance and reduce computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If power difference or correlation based double talk detection is used, then detection simplicity is maintained, but detection accuracy deteriorates when echo level exceeds near-end signal level
Solution Approach 1:
The patent transitions from time-domain power difference detection to frequency-domain spectral correlation detection. By analyzing spectral correlations in frequency subbands rather than raw power differences in time domain, the system achieves accurate double talk detection even when echo level exceeds near-end signal level, resolving the contradiction between detection simplicity and accuracy.
Solution Approach 2:
The patent changes the detection parameter from power difference to spectral correlation coefficient. This parameter transformation enables the detector to distinguish between echo and near-end speech by exploiting their different spectral characteristics, maintaining high detection accuracy across varying echo and speech levels while preserving computational efficiency.
2Reliability
If adaptive filter size is increased to improve echo estimation, then echo cancellation performance improves, but DSP engine resource consumption increases
Solution Approach 1:
The patent applies partial action by using a smaller adaptive filter than the full acoustic echo path length. The frequency-domain double talk detection mechanism compensates for the reduced filter size by detecting spectral correlations, allowing acceptable echo cancellation performance with lower DSP resource consumption.
Solution Approach 2:
The patent introduces frequency-domain spectral correlation analysis as an intermediary mechanism. This intermediary detection approach enables accurate double talk detection without requiring a large adaptive filter, thus bridging the gap between limited filter size and acceptable echo cancellation performance while conserving DSP resources.
3Reliability
If NLP is activated to suppress echo residual, then echo quality improves, but near-end speech may be clipped during double talk periods
Solution Approach 1:
The patent implements feedback control through accurate frequency-domain double talk detection. The spectral correlation-based detector provides reliable feedback about double talk conditions, enabling the system to activate or deactivate NLP appropriately. This feedback mechanism prevents speech clipping during double talk while maintaining echo suppression during single-talk periods.
Solution Approach 2:
The patent replaces time-domain power-based detection with frequency-domain spectral correlation detection. This substitution provides more accurate and reliable double talk detection, enabling better control of NLP activation to prevent speech clipping while maintaining echo quality.
Data Source
AI summary
A method of detecting double talk condition in hands free communication devices is disclosed. In general, the method in accordance with the teachings of this invention detects double talk conditions based on inherent frequency response differences between the transducers used and acoustical effect on the spectrum of the returned echo signal. An input signal from a far-end talker and an input signal from the output from an echo canceler are received by the detector. K spectral subbands are created for each input signal. From this K subbands q subbands are selected based on inherent frequency differences between the far-end transducer and a near-end transducer. The spectral echo residual power is estimated at each subband. The estimated spectral echo power and the output signal from the echo canceler for a selected subband are compared to a predetermined threshold. Based on this comparison, it is determined whether double talk conditions exist based on the comparison.