Lots of noises are added to the unspoken parts and overall quality is not worse - files provides

[haoheliu]

Lots of noises are added to the unspoken parts and overall quality is not worse - files provides

[MohammedMehdiTBER]

Upon further investigation, it has been determined that there is an issue with clipping at the tail and audible noise present between silences and distortion of background vocals. While Descript's Studio Sound appears to be a suitable solution, it should be noted that it also has the potential to erase background vocals.

[haoheliu]

Thank you for your interest in our model. Many apologize if it does not work as expected.

Generally, my suggestion is to use mode 0 only. But at the same time, the open-sourced model is not enough for production use and may encounter many bad cases. I wish I can improve the model but currently, I'm working on a different research topic for my PhD. When I have a chance on this topic again I'll try my best to achieve a better open-sourced version.

[cweaver-logitech]

I'd like to add my own experiences using voicefixer and comparing it against other methods (such as Deepfilternet2). voicefixer can do some brilliant work, very impressive on tough problems. For example, it removed the non-stationary interference both outside of speech and overlapping with little effect on the quality of speech.

Original

voicerfixer

At other times is adds lots of stationary noise to a seemly clean audio file.

I realise this is just a part project but, in broad terms, what would it take it improve the performance? Is it a question of a larger training dataset or it's it something more complex?