| license: mit | |
| tags: | |
| - audio-generation | |
| library_name: diffusers | |
| base_model: harmonai/jmann-small-190k | |
| Blunstron is a model I made for Harmonai's Dance Diffusion. | |
| The dataset is less than five minutes of the song Old and Wise by The Alan Parsons Project, yet it performs very well and does not overfit. | |
| Old and Wise is sung by Colin Blunstone, hence the name Blunstron. | |
| # Why | |
| I put music out for free on YouTube containing lots of tiny samples (in imitation of a musician named Todd Edwards), | |
| and I've sampled Old and Wise a LOT because I like the auditory textures in it. I'm kind of running out of potential chops, | |
| so I decided to generate a practically infinite supply of them with AI. | |
| # How | |
| I finetuned this on Google Colab for around two hours. I kind of dislike how the word "finetune" is used for Dance Diffusion, since unlike | |
| Dreambooth with Stable Diffusion, Dance Diffusion models (including this one) effectively | |
| become an entirely different model when fine-tuned for long enough. | |
| # Audio Characteristics | |
| - Deep, ethereal, soft auditory texture | |
| - Mostly chords, not much melody | |
| - Colin Blunstone-like vocals | |
| - Occasional drum hits | |
| A few examples are provided in the files of this git repo called blunstron-test-x.wav. |