Velvet Noise Reverberation  – demo page

This page contains demonstrations related to a paper submitted to the AES30 conference in Saariselkä, Finland, in March 2007:

Matti Karjalainen and Hanna Järveläinen

Reverberation Modeling Using Velvet Noise

Abstract:

Room reverberation consists of a multitude of reflections from surfaces and objects in a room. Particularly the late reverberation tail resembles noise with an exponential decay envelope. Artificial reverberation algorithms try to simulate this in a computationally efficient manner. Some proposed algorithms are based on the convolution with a sparse FIR filter corresponding to a randomized sparse sequence of unit impulses. In this paper we search for such sequences with minimal impulse density vs. maximal smoothness of the noise-like characteristics. Such noise is called here "velvet noise'', because it can sound smoother than the Gaussian noise. The perceptual characteristics of velvet noise are described by results from listening experiments and auditory analysis. Reverb algorithms based on velvet noise are discussed and analyzed. 

Demos: (click to play sound):

Full band noise samples:

  1. Reference signal: Gaussian white noise (full-band)
  2. Velvet noise, impulse density imp/sec: 7000, 3000, 1500, 1000, 800, 600
  3. RASI noise, impulse density imp/sec: 7000, 3000, 1500, 1000, 800, 600

1500 Hz lowpass noise samples:

  1. Reference signal: Gaussian white noise (full-band)
  2. Velvet noise, impulse density imp/sec: 7000, 3000, 1500, 1000, 800, 600
  3. RASI noise, impulse density imp/sec: 7000, 3000, 1500, 1000, 800, 600

Modeling of a concet hall (Pori, Finland, data ’S1R3o.m’):

150-tap SFIR is used for early sound (in addition to LPC for direct sound) and 2*75-tap 50ms recursive SFIR for late reverberation.
  1. Room impulse response: original vs. modeled
  2. Speech response: original (convolved) vs. modeled
  3. Sine wave response: original (convolved) vs. modeled

Notice the main problem with steady-state sinewave, where the 'modulation noise' and level growth are clearly audible.

This URL: http://www.acoustics.hut.fi/demos/VelvetReverb
Last modified: March 13, 2007
Author: Matti Karjalainen