1.2 Balance Dataset

[hughmancoder]

• Curate and expand the dataset to address class imbalance and limitations.
• More percussion recordings needed
• Include instrument classes in the project Scope, each with > 600 3s recordings
• Suona — double-reed horn
• Erhu — two-string bowed fiddle
• Pipa — four-string plucked lute
• Dizi — bamboo flute
• Guzheng — plucked zither
• Sheng — mouth-blown free-reed pipe organ
• Percussion — gongs, drums, and cymbals

Datasets

• Get CTIS and ChMusic datasets and integrate them into dataset. We can add the files to our instrument_name.json file

Dataset improvements

• Use data augmentation (mix in Chinese instruments with other instruments) to simulate realistic polyphony.

• We can do this by taking existing labels and overlaying audio signals

• Consider multi-label detection (presence/absence) rather than strict “which one dominant”.