Spaces:

fabiogra

/

moseca

like 52

Running

Dear Hugging Face team,

I am writing to apply for a community GPU grant for my open-source project Music Source Splitter.

Music Source Splitter is a web app built using Streamlit that allows users to separate the vocals and the instrumental of a song using machine learning. The code for this project is open sourced on both GitHub and Hugging Face.

This app uses a pretrained model called Hybrid Spectrogram and Waveform Source Separation from facebook/htdemucs. A GPU grant from Hugging Face would allow me to significantly improve this app and better serve in two ways:

1. Speed up the machine learning model behind this app. Currently, the app is limited to processing 10 seconds of a song at a time due to the model running on a CPU. A GPU would allow me to extend the processing time to the full length of a song and increase the speed of processing.

2. Improve the quality of separation by using a more advanced architecture of the htdemucs model. The latest versions of this model (v4) seems require more intensive GPU processing to generate higher quality outputs. Futhermore, it would allows me to serve htdemucs_6s that separate 6 sources version (piano and guitar in advance).

Thanks to this tweet from a streamlit developer advocate, the app just receive a bit of popularity and I decided to develop a second version.

Please let me know if you need any additional information. Thank you for your consideration.

Regards,
Fabio Grasso