Impulse-Response and CAD-Model-Based Physical Modeling in FAUST

International audienceCONTEXT : The FAUST programming language has proven to be well suited to implement physical models of music instruments using waveguides and model synthesis. We developed two tools allowing to easily generate FAUST modal physical models: 1. ir2dsp.py takes the audio file of an impulse response and converts it into a FAUST program implementing the corresponding modal physical model; 2. mesh2dsp.py outputs the same type of model from a .stl file specifying a 3D object. FAUST MODAL PHYSICAL MODEL : Linear percussion instruments can be implemented using banks of resonant bandpass filters. Each filter implements one mode of the system and is configured with 3 parameters : the frequency of the mode, its gain and its resonance duration (t60). Its FAUST version, modeFilter below, uses a biquad filter (tf2) and computes its poles and zeroes for a given frequency and t60. Modal physical models are implemented using multiple parallel (par in FAUST) instances of mode calls. The FAUST-generated block diagram corresponding to such an implementation is presented below (we used arbitrary parameters here). mode(100)(0.9f)(0.9f) mode(200)(0.8f)(0.9f) mode(300)(0.6f)(0.5f) mode(400)(0.5f)(0.6f) process Such a model can be excited by a filtered noise impulse. White Noise Lowpass Highpass Envelope To ModelIR2DSP.PY AND MESH2FAUST: ir2dsp.py takes an audio file and extracts modal physical model-based information for each mode: frequency and gain, by carrying out peak detection; t60, by measuring bandwidth at-3 dB. A FAUST file is then generated. With this tool, one can strike any object, record the resulting sound and turn it into a playable digital instrument. mesh2dsp.py gives the same output, using a .stl file (describing a 3D object) as input, as follows: • conversion of the input object to a mesh; • Finite Element Analysis (FEA) using the Elmer package, with the Young modulus, Poisson coefficient and density of the material as parameters: • frequency and gain computation from eigenvalues and mass participation for each mode; • t60 values input (these values cannot be computed by this method unfortunately, so they are user-provided parameters). EVALUATION : Spectrograms of (a) the recording of the IR of a can and (b) its ir2dsp.py-generated modal physical model: (a) (b) The original and synthesized sound representations are relatively close (but see Future Directions).FUTURE DIRECTIONS We plan to improve ir2dsp.py by using a better t60 measurement algorithm. For now, the calculation is done by measuring the band-width for each peak, while it would be a better approach to extract it from a time-frequency representation of the signal. Regarding mesh2dsp.py, we would like to try other open-source packages than Elmer to carry out FEA