AAI_2025_Capstone_Chronicles_Combined

6

or no category, and brightness is represented without regard to potential spectral changes throughout the audio.

For this project, NSynth provides a consistently labeled dataset that is suitable for training and evaluating both classical and learned timbre representations. Figure 1 illustrates an example of NSynth’s metadata features.

Figure 1

NSynth’s metadata features, data type, and description

Note: the “audio” feature is omitted from the JSON example since the audio data is stored separately in the WAV files keyed by “note_str”.

338