Digital Audio Signal Restoration Using Noise Estimation Algorithm

1
Digital Audio Signal RestorationUsing Noise
Estimation Algorithm

• May 19, 2006
• Sang Ki Park
• Electrical and Computer Engineering
• The University of Tennessee, Knoxville

2
Agenda

• Introduction
• Motivation
• Purpose of Project
• Noise Reduction Methods
• Short Time Fourier Transform
• Spectral Subtraction
• Noise Estimation Algorithms
• Noise Estimation Algorithms for Spectral
  Subtraction
• Experimental Results
• Conclusions

3
Motivation

• Restore Digital Audio Signal from Unknown Noise
• Noise Reduction with Efficient Noise Estimation

4
Purpose of Project

• Review Noise Reduction Method
• Survey Noise Estimation Algorithm
• Find Practical Methods for restoring real noisy
  audio

5
Noise Restoration

• Noise Model
• Assume noisy signal has zero mean
• Suppose clean signal and noise are uncorrelated
• Restoration model

6
Short Time Fourier Transform (STFT)

• Short Time Fourier Transform
• STFT divides input signal into small time frames
• STFT Noise Restoration

Noise spectrum estimate
Noise suppression rule
Spectral attenuation
STFT synthesis
STFT analysis
7
Noise Whitening

• Noise Whitening Filter
• Additive Noise has non-flat amplitude spectrum
• DC and low frequency bias give trend to original
  signal
• Noise Reference Signal Power Spectrum Density
• Noise Whitening filter is
  multiplied to noisy signal spectrum

8
Spectral Subtraction

• Noise Signal Model
• Power Subtraction
• Magnitude Subtraction

9
Spectral Subtraction (cont)

• Spectral Subtraction in STFT

10
Wiener Filtering

• Standard noise reduction technique
• Noise Reduction Performance Comparison
• Magnitude Subtraction is better than Power
  subtraction and Wiener filter

P. J. Wolfe and S. J. Godsil, Perceptually
Motivated Approaches to Music Restoration,
Journal of New Music Research, Vol. 30, No.1,
pp.83-92, 2001.
11
Noise Gate

• Suppress parts of the signal spectrum below the
  defined threshold
• attenuation ratio will be adjusted by changing
  the transform slope under the threshold level
• Remove Residual Noise

12
Noise Estimation

• Provides Noise information to Noise Reduction
  Process
• Noise has non-stationary character
• Noise reference is needed to be updated
  continuously
• Restored audio quality will depends on the noise
  estimation.
• Inappropriate noise estimation makes Improper
  noise reduction
• Improper noise reduction causes artifact sound,
  loss of important sound signal or no noise
  reduction
• Noise Estimation for Spectral Subtraction
• Noise level detect by using Low frequency
• Noise level detect by using High frequency
• Noise reference update by signal activity
  detector
• Noise reference update by noise tracker with SNR
  comparison

13
Noise Estimation using Low frequency

• Low-frequency 0-50 Hz contain no human vocal
  information
• Reference noise spectrum from
  No-signal (music) frame
• Adaptively Change Noise level in each frame
• Estimate Noise Restoration

14
Noise Estimation using Low frequency

• Block Diagram of Noise Estimation using Low
  Frequency

15
Noise Estimation using High frequency

• Human vocal information mostly exists between 50
  Hz and 3.5 kHz
• For finding noise level, Use high frequency more
  than 10 kHz
• Assume noise level from high frequency is same to
  that of low frequency
• Use Reference Noise Spectrum from noise only part
• Requires a very high sampling rate more than 20
  kHz

16
Noise Estimation with signal activity detector

• Signal Estimation based on spectral subtraction
• A local minimum noise power tracker
• Smooth in frequency
• Smooth in time
• Minimum noise tracker

17
Noise Estimation with signal activity detector
(cont)

• Signal Activity Detector
• If detect no signal activity then update noise
  reference

Signal presence
Signal presence probability update
Signal absence
18
Adaptive Noise Estimation with noise frame tracker

• Find the noise only frame and update noise
  reference
• Signal-to-Noise Ratios (SNR) are consecutively
  computed in three frequency bands
• (Low from 0 to 1 kHz, Middle over 1 kHz
  to 3 kHz, High over 3 kHz)
• Threshold is defined empirically
• If all three SNR are smaller than thresholds, the
  current frame is noise only frame

19
Adaptive Noise Estimation combined with noise
Spectral Subtraction
Start
Update frame number
Obtain Current frame Signal Spectrum
Calculate SNR in three Frequency bins
Yes
If
No
Noise reference update
Noise subtraction
is Last frame ?
No
Yes
End
20
Original Test Data

• Carusos Vocal with Piano Accompaniment
• Recorded in1920s, digitalized by 8kHz Sampling,
• Part of the song is target signal (50000 sample,
  6.25 second play time)
  

21
Manual Noise Reference Find

• Find Noise Only Part (Music Absence) Manually
• Noise only part
  Selected Noise Signal

22
Noise Whitening

• Make Noise Reference from Noise Only Part
• Noise Whitening Filter from the Noise Reference
• Noise Spectrum
  Smoothed spectrum
  Inversed spectrum

23
Noise Whitening

• Spectrum Comparison
• Original
  Noise Whitened

24
Noise Whitening

• Spectrogram Comparison
• Original
  Noise Whitened

25
Noise Whitening

• Noise Whitened Signal

Original noisy Sound
Noise Whitened
26
Noise Reduction

• Find Noise Reference for Reduction
• Noise Whitened Signal is given as input
• Take noise reference from the same part used for
  noise whitening
• Compare Noise Power Subtraction and Noise
  Magnitude Subtraction
• Manually Found Noise Spectrum

27
Compare Noise Reduction Methods (Spectrum)

• Spectrums are similar to each other
• Power subtraction
  Magnitude Subtraction

28
Compare Noise Reduction Methods (Spectrogram)

• Magnitude Subtraction remove noise more clearly
• Power subtraction
  Magnitude Subtraction

29
Compare Noise Reduction Methods (Signal)

• Magnitude Subtraction gave more precise result
• Power subtraction
  Magnitude Subtraction

30
Residual Noise Reduction

• Find Noise Gate threshold value as -40dB
• Magnitude Subtraction result is given as input
• Average Ceiling -25.13dB , Average Floor -54.43dB
  and center -39.78

31
Residual Noise Reduction

• Noise Gate is applied to Manual Noise Subtraction
  Result
• Threshold level is -40 dB
• Spectrum
  Spectrogram

32
Manual Noise Reduction and Noise Gate

• Restored Signal

Original noisy Sound
Manually Restored Sound
33
Adaptive Noise Estimation

• Track noise frame and update noise reference
• Original noisy Caruso song is given as input
• Frame size 1000, Total number of frame is 50
• Compare SNR continuously in High, Middle and Low
  Frequency

34
Estimated Noise Subtraction

• Subtract Estimated Noise Magnitude in each frame
• Spectrum
  Spectrogram

35
Comparison of Manual and Automatic Reduction
Manual Reduction
Estimated Noise Subtraction
36
Residual Noise Reduction

• Noise Gate is applied to the Estimated Noise
  Subtraction Result
• Threshold level is -40 dB

37
Estimated Noise Subtraction and Noise Gate

• Restored Signal

Proposed method
Original noisy Sound
38
Compare with Commercial Software

• Cool Edit Pro (Digital sound editor).

• popular audio tool
• mastering and analysis tools
• audio restoration
• Echo, Reverb, Flanging, Chorusing
• Compression, Limiting, Equalization
• Noise Reduction
• Clip Restoration

Noise Reduction Setting Window
39
Spectrum Comparison
Original noisy Sound
Proposed method
Cool Edit Pro
40
Spectrogram Comparison
Proposed method
Cool Edit Pro
41
Restored Sound Comparison
Original noisy Sound
Proposed method
Cool Edit Pro
42
Conclusions

• Review Noise Reduction method
• Survey Noise Estimation method
• Combine the Noise Estimation and Reduction
  methods

43
Reference

• S. Rangachari, P. C. Loizou and Y. Hu, A Noise
  Estimation Algorithm With Rapid Adaptation For
  Highly Non-Stationary Environments, Proc. IEEE
  Int. Conf. Acoustics, Speech, and Signal
  Processing (ICASSP 04), Vol. 1, pp.I305-I308,
  May 2004.
• S. Rangachari and P. C. Loizou, A
  noise-estimation algorithm for highly
  non-stationary environments, ELSEVIER Speech
  Communication 48 pp.220-231, Feb. 2006.
• S. L. Gay and J. Benesty, Acoustic Signal
  Processing For Telecommunication, Kluwer Academic
  Publishers, Massachusetts, 2000.
• P. J. Wolfe and S. J. Godsil, Perceptually
  Motivated Approaches to Music Restoration,
  Journal of New Music Research, Vol. 30, No.1,
  pp.83-92, 2001.
• http//sound.eti.pg.gda.pl/denoise/noisegate.html
• J. G. Proakis and D. G. Manolais, Digital Signal
  Processing Principles, Algorithms, and
  Application, third edition, Prentice Hall, New
  Jersey, 1996.

44
Reference

• M. Kahrs and K. Brandenburg, Applications of
  Digital Signal Processing to Audio and Acoustics,
  Kluwer Academic Publishers, London, 1998.
• S. M. McOlash and R. J. Niederjohn, A Spectral
  Subtraction Method for the Enhancement of Speech
  Corrupted by Non-White, Non-Stationary Noise,
  Proc. The IEEE Industrial Electronics, Control
  and Instrumentation Int. Conf. (IECON 1995), Vol.
  2, pp. 872-877, Nov. 1995.
• M. Schwab, H. G. Kim and P. Noll, Robust Noise
  Estimation Applied to Different Speech
  Estimators, 37th Asilornar Conf. on Signal,
  Systems and Computers (ACSSC), Vol. 2, pp.
  1904-1907, Nov. 2003.
• J. Yamauchi and T. Shimamura, Noise Estimation
  Using High Frequency regions for speech
  Enhancement in Low SNR Environments, IEEE
  Worshop Proc. Speech Coding, pp.59-61, Oct. 2002.
• K. Yamashita and T. Shimamura, Nonstatinoary
  Noise Estimation Using Low-Frequency Regions for
  Spectral Subtraction, IEEE Signal Processing
  Letters, Vol.12, Issue 6, pp. 465-468, June 2005.