Suppression of Musical Noise Artifacts in Audio Noise Reduction by Adaptive 2D filtering - PowerPoint PPT Presentation

Loading...

PPT – Suppression of Musical Noise Artifacts in Audio Noise Reduction by Adaptive 2D filtering PowerPoint presentation | free to download - id: 857683-ZGY4Z



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Suppression of Musical Noise Artifacts in Audio Noise Reduction by Adaptive 2D filtering

Description:

Title: Suppression of Musical Noise Artifacts in Audio Noise Reduction by Adaptive 2D Filtering Author: Alexey Lukin, Jeremy Todd Last modified by – PowerPoint PPT presentation

Number of Views:22
Avg rating:3.0/5.0
Slides: 19
Provided by: Alexe99
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Suppression of Musical Noise Artifacts in Audio Noise Reduction by Adaptive 2D filtering


1
Suppression of Musical Noise Artifactsin Audio
Noise Reductionby Adaptive 2D filtering
Convention paper 7168
Alexey Lukin AES Member Moscow State University,
Moscow, Russia
Jeremy Todd AES Member iZotope Inc., Cambridge, MA
2
Spectral subtraction
  • Reduction of additive stationary noise
  • Magnitude spectra are subtracted
  • Phase spectrum is left intact

spectral subtraction workflow
3
Musical noise artifact
  • Variance of noise spectrum leads to spurious
    bursts of energy after spectral subtraction

STFT of white noise, close-up and overall image
4
Musical noise artifact
  • Variance of noise spectrum leads to spurious
    bursts of energy after spectral subtraction

STFT of white noise before and after a simple
spectral subtraction
5
Musical noise artifact
  • Simple approaches for reduction
  • Oversubtraction
  • Loss of signal details
  • Mixing in the original noise
  • Only moderate noise reduction amount
  • Time smoothing of gates gain
  • Smearing of transients, noise echoes




6
Existing approaches
  • Use of time-smoothed energy estimates
  • Introduction of attack and release time for
    sub-band gates
  • Ephraim-Malah method combination of a-priori
    (time-smoothed) and a-posteriori
    (instantaneous) energy estimates

7
Existing approaches
  • Use of 2D filtering of a spectrogram
  • Whipple 1994 Explicit detection and elimination
    of energy bursts
  • Goh et al 1998 Detection of musical noise by
    analysis of local variance and use of median
    filter for its elimination
  • Lin/Gourban 2003 2D smoothing for signal
    detection, 1D time smoothing for processing
  • Soon/Koh 2003 1D DFT applied to rows of a
    complex spectrogram, coefficient shrinkage

8
Non-Local Means algorithm
  • Recently proposed NLM algorithm Buades 2005
  • Image denoising by comparison of patches

9
Non-Local Means algorithm
  • Non-Local Meansfor image denoising
  • Capable of preserving and enhancing the 2D
    structure

Weights are high for q1, q2, but not for q3
Illustration from Buades et al 2005
10
Application to audio
  • Use NLM algorithm to smooth SNRf,t
  • Spectrograms have a prominent 2-dimensional
    structure harmonic peaks, frequency-correlated
    transients
  • NLM, unlike 1-dimensional time smoothing methods,
    is able to account for frequency correlations in
    a spectrogram

11
DFT thresholding
  • DFT thresholding applied to spectrogram patches
    (again, similarly to image denoising)
  • Sparse representation of harmonics ? better
    signal/noise separation
  • Different type of artifacts compared to NLM ? a
    hybrid algorithm will have less artifacts of each
    type

12
Results
Spectrogram of a noisy signal SNR 15 dB
13
Results
After simple spectral subtraction SNR 21.7 dB
14
Results
Result of Ephraim-Malah suppression SNR 21.0 dB
15
Results
Result of the proposed method SNR 21.8 dB
16
Results
  • Features of the proposed method
  • Adjustable amount of suppression of musical noise
  • Adaptive 2D smoothing minimizes loss of signal
    details
  • Computationally expensive (but allows effective
    parallel implementation on multi-core processors)



17
Results
  • Improvement of SNR after different methods

Method \ SNR 25 dB 15 dB 5 dB
Simple spectral subtraction 4.44 6.69 9.74
Ephraim-Malah method 3.96 5.98 9.46
The proposed method 4.53 6.79 9.98
18
Your questions
?
Demo web page http//www.izotope.com/tech/aes_sup
pr/
About PowerShow.com