Isolate Vocals for FREE!

[basilwoods]

Hi, Ive created a website that uses AI (spleeter) to isolate vocals and instrumentals on any song. www.audiostrip.co.uk

It is completely free so worth a try. Let me know what you think

Cheers

[James Perrett]

That's a very interesting website. I hadn't heard of Spleeter. I tried one of the files that I had previously used with Izotope's Music Rebalance and the results from your website were very similar. I would be interested in trying it with a more challenging file but, as that file belongs to a client, I would currently be worried about uploading that file to an unknown website which, for all I know, could be storing copies of everything.

[James Perrett]

Though I've just done a little more reading and it appears that Spleeter only works up to 11kHz - which was fine for my compressed test file but wouldn't be any good for more serious work.

[basilwoods]

Though I've just done a little more reading and it appears that Spleeter only works up to 11kHz - which was fine for my compressed test file but wouldn't be any good for more serious work.

Ok thanks for trying it anyways. I believe it is possible to get Spleeter to go up to 16kHz if that makes it better? Although I have not tested at that range so may work to varying degrees of success.

Also I store data only long enough for you to make sure youve downloaded it if needed (all files are deleted after 3 days)

Cheers

[James Perrett]

Thanks for the response Basil. It still looks a useful tool, even with the bandwidth limitation.

[VOLOVIA]

Hi, I have tried on one of my songs, which of course I know the sound of the individual tracks...
The new backing track sounds cool, with just ghosts of the vocals. I suppose, it would be good enough to replace the vocals and masking the leftovers of the old ones.

The extracted vocal stem was not too bad either. I suppose with some careful editing (it picked up the solo guitar in places...) it could be used in (busy) club remixes.

Good and useful attemp in any case! Cheers.

[basilwoods]

Hi, I have tried on one of my songs, which of course I know the sound of the individual tracks...
The new backing track sounds cool, with just ghosts of the vocals. I suppose, it would be good enough to replace the vocals and masking the left overs of the old ones.

The extracted vocal stem was not too bad either. I suppose with some careful editing (it picked up the solo guitar in places...) it could be used in (busy) club remixes.

Good and useful attemp in any case! Cheers.

Thanks for trying it and thanks for the feedback

[James Perrett]

I'm just bumping this because I've been doing a little research into the source separation options that are currently available. I've had a few projects recently where RX's Music Rebalance has certainly earned its keep but I have also just demo'd Acon Acoustica for one project because it can also separate piano though I'm not sure that I can justify the purchase price for just one or two songs. So I did a search for alternatives and it appears that Basil's website is keeping up with the current fairly rapid progress. So, if you've tried

https://www.audiostrip.co.uk/

in the past it would be worth revisiting it and trying some of the new options.

In particular, the Demucs V3 option offers far better sound quality than the Spleeter option. This makes the Music Rebalance feature of Izotope's RX pretty much redundant as it produces results that are nowhere near as good as Demucs V3. The only reason for using Spleeter (and its derivatives like RX) is for the ability to separate piano as well as vocals, drums, and bass. Since RX doesn't offer the piano separation option you would need to go to something like Acon Acoustica if you need that feature in a paid program but Audiostrip offers it for free.

[James Perrett]

This whole field is moving ahead rapidly and I thought I would add to this thread to mention that Demucs V4 is now available (and supported by the Audiostrip and MVSep websites). I'm just in the process of checking it out now but there are a couple of new models available.

The first is the Hybrid Transformer method

The v4 version features Hybrid Transformer Demucs, a hybrid spectrogram/waveform separation model using Transformers. It is based on Hybrid Demucs (also provided in this repo) with the innermost layers are replaced by a cross-domain Transformer Encoder. This Transformer uses self-attention within each domain, and cross-attention across domains. The model achieves a SDR of 9.00 dB on the MUSDB HQ test set. Moreover, when using sparse attention kernels to extend its receptive field and per source fine-tuning, we achieve state-of-the-art 9.20 dB of SDR

The second is an extension of the Hybrid Transformer model which separates a file into 6 separate sounds - drums, bass, other, vocals, piano and guitar. Apparently it is all a bit experimental at the moment but I'll add to this thread when I've had more of a play.

It is also worth mentioning that there are a couple of Dumucs GUI programs for Mac and Windows so you no longer need to manually install things from the command line.

https://carlgao4.github.io/demucs-gui/

and

https://github.com/Anjok07/ultimatevocalremovergui

I've used the first one but it doesn't currently support Demucs V4. The results seem identical to the Google Colab version that I've used previously. I've just downloaded the second (which does support Demucs V4) to see how it compares.

[Tomás Mulcahy]

It's amazing tech. Worked great last year for a friend/ client who needed to redo vocals but no longer had multitracks. Audiostrip was much better than the Izotope implementation.

I've still not had success with extracting piano. I wonder if this is because all the stuff I've tried is me playing digital piano live (rack sampler) rather than real piano, or perhaps it is because the source is either cassette or minidisc? I just get warbled useless stuff or nothing at all. Anyway, I'm sure by this time next year it will be possible.

[James Perrett]

I've still not had success with extracting piano. I wonder if this is because all the stuff I've tried is me playing digital piano live (rack sampler) rather than real piano, or perhaps it is because the source is either cassette or minidisc? I just get warbled useless stuff or nothing at all. Anyway, I'm sure by this time next year it will be possible.

I've just tried the Demucs version that can extract piano with mixed results. On one song it hardly recognised the piano at all - possibly because there was quite a bit of chorus and reverb on the piano sound. In contrast, Spleeter did recognise the piano and mainly extracted it successfully. However, as with most spleeter separations, there was a slight warbling.

But all was not lost - Demucs did recognise the guitar and, as this track was piano, bass, drums and guitar, the piano was mainly isolated on the "other" track with just a few notes on the piano track.

On another song with more natural sounding piano bass and drums, the piano was isolated successfully although it sounded heavily gated with distortion, probably due to the gating. The bass was also slightly distorted which wasn't an artefact that I had noticed with the older Demucs algorithm.

All this was tested using the UVR software from the second link in my previous message which seems to be a very handy way to use Demucs.

[Tomás Mulcahy]

Interesting. I will try Spleeter, Basil’s site runs that also. I only tried it once at it didn’t work well.

Another useful tool is Fuse Audio Labs drums SSX. Used on stems like we get from Demucs it’s useful for tweaking the balance of the kit. But only tweaking, not outright unmixing. For that I’ve successfully used it to extract accurate triggers, although it gets very time consuming to recreate tom fills.