RNN Parameterization

This page is a companion page for the paper:

Wyse, Lonce. (2018), Real-valued parametric conditioning of an RNN for real-time interactive sound synthesis. 6th International Workshop on Musical Metacreation, International Conference on Computational Creativity (ICCC) June 25-26, 2018, Salamanca, España

PyTorch code coming soon...
Figure 4 :

Figure 4. A network was trained only on the extreme low and high pitches at the endpoints of the parameter value range. During generation, the parameter value was swept through untrained values between its lowest and its highest and back again over 3 seconds. The result is this continuously, though non-linearly varying pitch.

Figure 5 :

Figure 5. The trained "Synth even" instrument controlled with an arpeggio over the pitch range illustrates the model's ability to respond quickly to pitch changes. This image also shows that the untrained middle pitch values are not synthesized as clearly as the trained values at the extremities. Furthermore, the middle values contain both even and odd harmonics, thus combining timbre from each of the two trained instruments.

Figure 6 :

Figure 6. When the network was trained on real data with only two extreme values of pitch, pitch had a more pronounced "stickiness" to extreme trained pitch values showing a transition region without a glide as the pitch parameter moves smoothly from low to high and back.

Figure 7 :



Figure 7. a. The clarinet, trained on 13 chromatic notes across an octave, generating a sweep with the pitch parameter swept from low to high and back. b. The trumpet playing the arpeggio pattern. c. A continuous forth-andback sweep across the instrument parameter trained with the trumpet and clarinet at its endpoint values.