Timbre Audio Dataset (DX7-clone synthesizer)

22.5 hours of synthesized audio using the open-source learnfm clone of the DX7 FM synthesizer, based upon 31K presets from Bobby Blue. These represent ``natural'' synthesis sounds---i.e.presets devised by humans.

We generated 4-second samples playing midi note 69 (A440) with a note-on duration of 3 seconds. For each preset, we varied only the velocity, from 1--127, and perceptually normalized the level of each sound. Sounds that were completely identical were removed from the dataset. DX7 FM synthesis is good for this purpose because it doesn't have a noise oscillator. Thus, for a particular preset, there is a timbral variation as the velocity increases. 8K presets had only one unique sound. The median was 51 unique sound per preset, mean 41.9, stddev 27.4.

We used the Surge Python API to generate this dataset.

Applications of this dataset include:

• Timbre ranking within a preset
• Predict a sound's preset