Restoration and Enhancement of Instrumental
Recordings based on Sound Source Modeling
Paulo Esquef, Vesa Välimäki, and Matti Karjalainen
Sound Examples
1. Introduction
This page presents sound examples of audio restoration and enhancement based on Sound Source Modeling (SSM). We present a case based on the commuted waveguide synthesis algorithm for plucked strings, in particular, for single acoustic guitar tones. Comparisons with traditional approaches are also provided.
2. SSM and Enhancement
Enhancement here is related to extending the bandwidth of the signal. We start with a lowpass filtered version of the signal and use a SSM-based scheme to recreate the missing spectral information. The basic scheme consists of estimating suitable model parameters from the degraded signal, obtaining the corresponding excitation via inverse filtering, and using a processed version of the excitation to resynthesize a signal in which the lost information is restored.
Original: original.wav
Lowpass filtered version: lpf1000Hz.wav
Enhanced versions:
1) Noise burst with flat spectrum, SNR 20 dB: enhanced1.wav
2) Noise burst with colored spectrum, SNR 40 dB: enhanced2.wav
3) Noise burst with colored spectrum, SNR 20 dB: enhanced3.wav
4) Noise burst with colored spectrum, SNR 10 dB: enhanced4.wav
2. SSM and Audio De-hissing
.We are considering a case in which the signal to be analyzed is corrupted by zero mean Gaussian white noise.
Test Signals:
Original guitar tone: original.wav
Original + white noise (SNR=10dB): ogn10.wav
2.1. Signal Modeling Approach
Typical problems with spectral-based de-hissing methods are related to the trade-off between the musical noise and the smoothing effects perceived in the restored signal.
- Wiener Filtering (normal noise variance estimate): ogn10r1.wav
- Wiener Filtering (overestimated noise variance): ogn10r2.wav
In the first case, a residual noise is clearly perceived, while in the second the restored tone is too smoothed.
2.2 Source Modeling Approach
- Aggressive de-hissing the signal and then applying the bandwidth extension procedure
- Aggressively de-hissed (overestimated noise variance): ogn10adg30.wav
- Bandwidth extended after 1.: ogn10adg30be.wav
- Integrated de-hissing and bandwidth extension
- Aggressively de-hissed excitation (attack part preserved): ogn10ade.wav
This URL: http://www.acoustics.hut.fi/publications/papers/aes110-ssm/
Last modified: 19.02.2001
Author: <esquef@acoustics.hut.fi>