Page 23 - Spring2020

Figure 2. Left: instrumented clarinet mouthpiece for measurements with human players. Two pressure transducers are used to measure the pressure inside the instrument (mouthpiece pressure) and inside the player's mouth (blowing pressure). A strain gauge is used to monitor the vibrations of the reed. Right: measured and synthesized mouthpiece pressure for an excerpt from Weber's Clarinet Concerto no. 2 (see also Multimedia1 and Multimedia2 at acousticstoday.org/hawleymedia).
method and yields information regarding the pressure and flow inside the tube, whereas the radiated pressure and the reed state are also calculated.
To validate this player-instrument interaction model, the numerically synthesized signals have been compared to mea- surements using (1) an artificial blowing machine and (2) experiments carried out with human players. The blowing machine shown in Figure 1 establishes the conditions for repeatable laboratory measurements. The mouthpiece pres- sure obtained via the blowing machine is compared with the one calculated using a physical model for different articula- tion conditions (Chatziioannou et al., 2019). The success of this resynthesis process points toward the fact that all signifi- cant physical phenomena that take place during such note transitions are accurately captured by the physical model.
Figure 2 shows a similar comparison between human play- ers and physics-based modeling. The blowing pressure and the mouthpiece pressure are measured during a performance using two pressure transducers (one inside the player’s mouth and one inside the mouthpiece), whereas the reed displace- ment is monitored by means of a strain-gauge sensor. The model is capable of qualitatively reproducing an excerpt per- formed by a professional musician (taken from the Clarinet
Concerto no. 2 by Carl Maria von Weber; see Multimedia1 and Multimedia2 at acousticstoday.org/hawleymedia).
During the above inverse-modeling applications, the fact that the involved model parameters have a physical nature allows intuitive access to them in terms of both control and interpretation (Campbell et al., 2004). Hence, it is possible to analyze the function of an instrument as well as the control exerted on it by the player by studying how the parameters of the physical model vary during a performance. Such physics- based analysis requires the formulation of accurate models and the design of suitable experimental setups to deter- mine those parameters that influence the sound generation mechanism. Deep neural networks have also been used for parameter estimation and sound resynthesis (Gabrielli et al., 2018), but their application has been limited on isolated notes and yielded signals that are only qualitatively similar to the recorded ones. Such networks can be used for more than parameter estimation; in Neural Audio Synthesis, we discuss synthesizing waveforms directly using neural networks.
Neural Audio Synthesis
In contrast to physics-based modeling, NAS approaches seek to minimize the difference between a recorded audio sound (or set of sounds) from a musical instrument and audio
Spring 2020 | Acoustics Today | 23