Download presentation

Presentation is loading. Please wait.

Published byBryan Howard Modified over 3 years ago

1

2
Real-Time Bayesian GSM Buzz Noise Removal Han Lin and Simon Godsill University of Cambridge Signal Processing Group

3
Outline Introduction to GSM Buzz Noise Pulse and the Restoration Model Detection of Noise Pulses Removal of Noise Pulses Audio Demo and Results Future Directions

4
What is GSM Buzz? Cellular phone (GSM,TDMA, and CDMA) send out strong electromagnetic (EM) pulses during registration process These pulses are received by audio amplifiers and line in circuits and causes noise known as GSM Buzz Buzz

5
GSM Buzz Identification Visual representation of GSM Buzz GSM Buzz (Interference Pulses) Audio representation of GSM Buzz GSM Buzz can be everywhere

6
Current Solutions to GSM Buzz Reducing cell-phone transmission power Changing transmission protocol Equipping a telecoil (hearing aid) Shielding All these solutions require hardware changes and are very difficult and expensive signal processing approach

7
Practical Applications AV/ PA equipments Recording studio Desktop and car stereos Portable players and recorders Telephones Hearing aids Statistical signal processing approach can provide last stage restoration for :

8
Analysis of Noise Pulse Central Pulse (constant width clock) Decaying Tail (capacitance) 217 Hz + harmonics

9
The Restoration Model x(n) - corrupted signal g(n) - known interference template b - constant scaling factor for amplitude difference e(n) - white output noise s(n) – original signal m - location of the start of the noise pulse

10
Design Strategy for GSM Buzz Removal Assume Interference Template is known (or can be measured) Assume central pulse has constant width Detect Noise Pulse location - m Estimate the scale factor - b Remove Noise Pulse one by one

11
Detection of Noise Pulses Hardware Electromagnetic wave detector Threshold detection/ slope detection Cross correlation/ matched filter Bayesian step detector Autoregressive detector The Bayesian template detector Detect Detection is generally not a problem

12
The Bayesian Template Detector x(n) - corrupted signal g(n) - known interference template s(n) – original signal, assume to be autoregressive b - constant scaling factor for amplitude difference m - location of the start of the noise pulse

13
The Bayesian Template Detector s(n) – original signal, assume to be autoregressive A contains AR coefficients a(i)

14
The Bayesian Template Detector Assume Where k is large constant We wish to integrate out parameters b and σ 1 in the detector to obtain an equation of only variable m Define probability model for The Bayesian template detector :

15
The Bayesian Template Detector Solution for The Bayesian template detector :

16
Performance of Bayesian Template Detector Interfered Signal Bayesian Template Detector Plot P(m|x,g) MAX P(m|x,g) m

17
Removal of Noise Pulses with AR Template Interpolator LSAR interpolates the data in the central pulse region (assume data missing) Iterative model: s(n) – original signal, assume to be autoregressive x(n) - corrupted signal g(n) - known interference template b - constant scaling factor for amplitude difference m - location of the start of the noise pulse

18
Least Square AR Interpolator LSAR interpolates the data in the central pulse region (assume data missing) Iterative model: Assume x is autoregressive Solve for a(i) and the solution for LSAR is:

19
AR Template Interpolator iterate r is estimated interference minimize e(n) to get b b Dotted : corrupted Green : original Red : estimate dip

20
Analysis of AR Template Interpolator Central pulse Decaying tail Green : original Red : first estimate Black: second estimate

21
GSM Debuzz Demo Interference Pattern Original Audio Interfered Audio Restored Audio

22
GSM Debuzz Demo ( Pop and Speech) Original Audio Interfered Audio Restored Audio PopSpeech

23
GSM Debuzz Results No audible artifacts and improve SNR by 50dB www-sigproc.eng.cam.ac.uk/~hl309/DAFX2006/

24
Real-time Consideration For detection, use threshold detector or hardware EM detector For restoration, use only one iteration LSAR interpolation has computation complexity of O(L^2) using levinson- Durbin recursion L is around 25 to 75 samples for CD quality audio

25
Future Works Exponential decay model Model the interference pulse as two exponential decays, estimate data in the central pulse region

26
Future Works Multi-channel Extension Model the noise pulse of one channel as a scaled version of the other channel Scale

27
Thank You

28
Real-Time Bayesian GSM Buzz Noise Removal Han Lin and Simon Godsill University of Cambridge Signal Processing Group

Similar presentations

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google