This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Synchronisation Errors?Synchronisation Errors?The performance of time-frequency
masking with respect to small phase and synchronisation errors is extremely robust.
Mixture:
• No error:
• Delayed by 1 sample:
• Delayed by 10 samples:
SNR improvementSNR improvement
Performance MeasuresPerformance Measures
SNR is a standard performance measureBut what about speech quality?Incorrect partitioning of t-f domain reduces
intelligibility of output.Introduce measure of WDO:
O. Yilmaz and S. Rickard, "Blind Separation of Speech Mixtures via Time-Frequency Masking", IEEE Transactions on Signal Processing, To appear, July 2004.
WDOWDO
MINUET Channel EstimateMINUET Channel EstimateFind set of t-f points, S, such that
for
otherwise
),(')(),( nHx
Adaptive TestingAdaptive Testing
Algorithm SNR (dB) WDO
NLMS 0.54 0.12
RLS 10.11 0.9
MINUET 15.18 0.76
Algorithm SNR (dB) WDO
NLMS -7.46 -4.57
RLS -1.36 -0.38
MINUET 7.69 0.44
Unity Channel:
Random Channel:
Conclusions and Future WorkConclusions and Future Work
MINUET estimates the channel and removes interference using instantaneous t-f magnitudes only.
This creates extraordinary robustness to phase errors when compared to classical adaptive filtering methods.
Improvements in t-f masking still necessary to increase intelligibility.
Algorithm complexity has not yet been considered. We presented pilot tests serving as proof of concept only. More realistic testing must be done to genuinely assess
performance. MINUET will be effective for any signals which are WDO.
Thank you for your attention!Thank you for your attention!