Analysis-Synthesis of Impact Sounds by Real-Time Dynamic Filtering

Authors: Aramaki M., Kronland-Martinet R.
Publication Date: March 2006
Journal: IEEE Transactions on Audio, Speech and Language Processing (Volume 14, number 2, pages 695-705)



This paper presents a sound synthesis model that reproduces impact sounds by taking into account both the perceptual and the physical aspects of the sound. For that, we used a subtractive method based on dynamic filtering of noisy input signals that simulates the damping of spectral components. The resulting sound contains the perceptual characteristics of an impact on a given material. Further, the addition of very few modal contributions—using additive or banded digital waveguide synthesis—together with a bandpass filtering taking into account the interaction with the excitator, allows realistic impact sounds to be synthesized. The synthesis parameters can be linked to a perceptual notion of material and geometry of the sounding object. To determine the synthesis parameters, we further address the problem of analysis-synthesis aiming at reconstructing a given impact sound. The physical parameters are extracted through a time-scale analysis of natural sounds. Examples are presented for sounds generated by impacting plates made of different materials and a piano soundboard.
Index Terms—Analysis-synthesis, audio virtual reality, banded digital waveguide, impact sounds, timbre, time-varying filtering

Synthesis of the material contribution (paragraph (III.A))

The main characteristics of an impacted material are reproduced with the model (without adding any modes).

Strong damping: the sound present some “wooden” characteristics

Weak damping: the sound present some “metallic” characteristics

Simulation from the two implementations of the synthesis model (see paragraph (III.D))

Simulations for an arbitrary set of parameters. The dynamic filter F is arbitrarily chosen as a one pole-one zero filter. The time evolution of the synthesis parameters are shown on Figure 3. The initial white noise is filtered by a bandpass filter to take into account the impact strength of the excitator. We arbitrarily chose to simulate five emergent modes for the two approaches (for the additive approach, a sum of five sinusoids and for the digital waveguide approach, five feedback loops). Signals are represented on Figure 4.

The sounds generated by the two approaches are similar. An important difference between these two approaches is that for the additive approach, the modes are obtained from a stationary signal (sinus waves) and for the banded digital waveguide approach, the modes are obtained from a non-stationary signal (white noise). Consequently, the sounds generated by the second approach seem more realistic because the deterministic part (contribution of the most prominent modes) of the signal is more correlated to the stochastic part.

To increase the realism of the sounds generated by the additive approach, we can replace the sum of sinusoids by a sum of filtered white noises centered around the frequencies of the sinusoids.

Example of synthesis signals, provided the same set of parameters:

For the second sound, the fluctuant characteristic of the filtered white noises allows the resonant part of the signal to be less separated from the noisy part.

Sound examples of impacted rectangular thin plates of various materials (see paragraph (IV.C.1))

The following signals were obtained experimentally by impacting rectangular thin plates.

Example of analysis-synthesis of an experimental signal (see paragraph (IV.C.2))

We have resynthesized the impulse response of a piano soundboard, obtained experimentally.
The synthesis parameters were estimated from the analysis of the original signal.

Control of the real-time synthesis model through the Radio Baton interface

The model has been implemented in real-time using the Radio Baton interface and for our purpose, two batons are used. With one baton, we control the evolution of the parameter to vary. With the second baton, we lauch the sound and control the strength of the impact.

Simulation of a morphing effect between distinct materials, simulating the continuous transition from the sound generated by a material to another one (from a metallic sound to a wooden sound):

Simulation of impact sounds on a metallic structure with various strengths of the impact. This example illustrates the control on the second baton:

Simulation of impact sounds on metallic structures of various dimensions. For a given set of frequency mode values, the variation of the dimensions is simulated by modifying the frequencies through a single multiplying factor:

Morphing effects, simulating the continuous transition between two materials

Each transition is composed of 22 consecutive hybrid sounds obtained by interpolating thesynthesis parameters. These sound examples are used in neurophysiological experiments to study the categorization process of impacted materials.