Wavetable-Inspired Artificial Neural Network Synthesis

We present WTIANNS, a method of database-driven sound synthesis with significant resemblance to wavetable synthesis. The primary difference between the technique presented and standard wavetable synthesis is that our method uses a multilayer perceptron (MLP) to generate wavetables in realtime. The MLP provides a continuous mapping from a set of psychoacoustic timbral features of an analyzed single-cycle waveform segment to the harmonic amplitudes of that waveform segment. In particular, we use either Bark spectrum or Bark Frequency Cepstral Coefficients (BFCCs) as input parameters. The results of the mapping are highly dependent on the character and timbral variance of the analyzed sound. Two significant advantages of our method over wavetable synthesis include its accommodation of higher dimensionality and the fact that the waveforms are automatically derived from a pre-existing recording. WTIANNS is not intended to emulate the recording it is based on, but to generate waveforms that are related–whether by interpolation or extrapolation–to those of the recording. This method encourages sound designers to explore the timbre space of recordings in a manner that is divorced from time.