Interpolation of Long Gaps in Audio
Signals Using Line Spectrum Pair Polynomials
Companion
webpage with sound examples of the homonym correspondence submitted to
the IEEE Transactions on Speech and Audio Processing
1. Original Test Signals
The test signals are high quality musical excerpts including a single
low-pitched piano tone and music genders such as pop, classical, and singing a
capella.
All signal excerpts were sampled at 44100 Hz and converted to .WAV files
(monaural, 16 bits, signed).
Classic: a 13-s-long fragment of orchestral music
Piano: a 3.7-s-long isolated low-pitched piano tone
Pop: a 14.3-s-long segment of Finnish pop music
Singing: a 20-s-long excerpt of singing a capella
2. Degraded Versions
Several gaps of 2000 samples were inserted in the signals by simply setting
the amplitude of the degraded portions to zero.
The gaps occur periodically in the signal. The gap advance is of 50000 samples.
3. Restored Versions
The degraded signals are restored through the original interpolation method presented in [10], [11], and [15] and via the proposed modified algorithm. For convenience we denote the conventional method by (Conv.), the proposed modified scheme by (Mod.).
For the original and modified interpolation schemes the length of the adjacent segments used to estimate the AR model is fixed at N = 2000 samples. The employed model orders p are indicated in the table below. The weighting factor used in the WLSP-based scheme is h = 1, which corresponds to replacing conventional AR models with their symmetric line spectrum polynomial.
|
Original |
Corrupted |
Restored |
||
|
|
|
Conv. p = 1000 |
Conv. p = 150 |
Mod. p = 150, h = 1 |
This URL: http://www.acoustics.hut.fi/publications/papers/tassp-int
Last modified: 23.12.2003
Author: <esquef@acoustics.hut.fi>