Skip to content

Latest commit

 

History

History
13 lines (9 loc) · 760 Bytes

README.md

File metadata and controls

13 lines (9 loc) · 760 Bytes

~GPT2 (181.25M) with Native Audio Generation... and Birds!

Chirps in, chirps out. The idea is for the model to be able to take a short recording and respond with an appropriate (similar) birdsong. With luck.

Inspired by https://github.com/nivibilla/build-nanogpt/tree/audio and Kate the Great.

Data: https://huggingface.co/datasets/DBD-research-group/BirdSet

WIP, plan is to pull from XCL, HSN, SNE and possibly POW & SSW recordings with quality A or B, latitude between 33 and 51 and longitude between -125 to -85, at least 6 seconds long and with at least (length // 6 seconds) entries in detected_events (and ultimately at least one event per 6 second chunk kept).

To train...

  • run python sherlock.py
  • run python train_gpt2.py

Samples