Suggestions, Questions, and Feature Requests for Fast Music Remover

[omeryusufyagci]

I'd love to get your feedback on Fast Music Remover, a tool being built truly with users like you in mind!

Here's some context on where we are and what we want to achieve in the near future:

We aim to provide an efficient core (MediaProcessor), that supports multiple ML models, and other modules, e.g. DSP techniques, that can address your needs and pain points in processing media files.

We've started with DeepFilterNet as our initial model, which helps us achieve efficient speech isolation. In the future you can expect to see support for more models, such as demucs, to provide a better coverage of potential use cases. These models and techniques all have their strengths and our vision for our core is to enable you with all the tools you may need!

We're focused on building an efficient core with a clean, modular API. Our goal is to make it easy for developers to integrate MediaProcessor into their projects. While the core currently exists in application form for CLI convenience and testing, the processing API is already designed to function as a library.
In the future, we plan to introduce a dedicated CLI tool and provide MediaProcessor as a standalone library for broader use. Stay tuned!

We're also developing a project launcher that will take care of installing dependencies, configuring project files with respect to your platform, and providing a streamlined way to access our core features. With the web application and docker image we're providing, you can seamlessly use Fast Music Remover on your local machine or remote servers!

Aiming to establish a stable release soon, we'll then take on realtime processing to support live feeds as well as the integration of other models and modules for our core.

If you have any use cases that we're not supporting at the moment, or have any feedback really, please don't hesitate to get in touch, here or on discord!

[passes0d]

Can this be used to extract the music into its own file? How does it work with music that contains vocals?

[omeryusufyagci]

Hi @passes0d! The main goal of this project is to clean media of noise, music, and effects. For this purpose, we're using DeepFilterNet, which isn't designed for extracting instrumentals. For extracting individual instruments, I recommend looking into demucs-based projects, as that model is more suited for this purpose.

Regarding music as input: you may need to adjust DeepFilterNet's parameters for filtering aggressiveness. Our default values are optimized for mainstream media, but they may not work as well for songs directly. I plan to introduce a higher-level setting for dynamically adjusting filter aggressiveness, but it's not available yet.

Thanks for your interest!

[rljacobson]

This feature request is probably out of scope for this project, but I would love for someone to develop a tool to "upscale" old recordings. How cool would it be if we could make old radio programs like Yours Truly, Johnny Dollar and audio from old films sound like modern studio recordings?

[omeryusufyagci]

Hi @rljacobson, thanks for bringing this up; it's always great to hear about interesting use cases from the community!

You're right, this would be something to explore after we've established our stable release. That said, I’d be open to supporting an upscaling module in the future. My vision for the core is to support various models and modules, to make it easy for users to choose the right features based on their needs. While we still have a long way to go, we’re making steady progress !

As our hands are full at the moment, I can't take this on immediately, but if there is interest from the community in contributing an upscaling module to MediaProcessor I'd be happy to support it as best as I can.

Thanks again for sharing this idea!

[eyewee]

Hello, are there any ways to remove just music (instrumental sound) and not the special effects ? (like animal noises, rain sound etc)?

[omeryusufyagci]

Hi @eyewee, thanks for getting in touch and your interest in our project !

The default parameters you're using now are selected for maximum attenuation. It might get a bit better if you were to use a less aggressive value for the filtering, and quite soon I will expose this on the UI so it can be easily selected by end users. The support is already available in the core though -in case you want to hack at it in the meantime...

So, currently, your best bet would be to play with filter_attenuation_limit and try to reduce it a bit to see if it helps.

The model we're using right now is intended to rapidly isolate vocals and it isn't really meant to perform this distinction between music and environmental sounds. This was a conscious choice, as we certainly want to support live feeds in the future. However, at the same time my vision is to support a set of models that would cater to different use cases. That will come with a performance overhead, and it will be up to the user to define their requirements.

To answer your question, it's not exactly here yet, but definitely something that's inline with the vision of the project. So, definitely stay tuned for updates! I hope this helps, and please don't hesitate to get in touch again! (we even have a discord if you prefer)

[HubKing]

I really truly do hate background music on YouTube videos and podcasts. Why do they keep adding loud music to their videos and audios? I haven't tested this library yet, but I did watch the demo, and it seems to be able to do what I want. Thanks in advance.

[omeryusufyagci]

Hey @HubKing, I'm glad to hear this will help you take control of the media you consume! Please let us know if you have any further feedback, and I hope you get some good use out of it! : ))