3

At the weekend I was listening to BBC Radio 2 and they played Eva Cassidy performing "Autumn Leaves" with the London Symphony Orchestra, which I really enjoyed.

As far as I know, the producers had a guitar-plus-vocal recording of the song, and somehow they managed to remove the guitar part entirely. I found a web page that says:

This version of the song was created by isolating Cassidy’s 1996 live vocal performance of the song, recorded at the Blues Alley Jazz Club in her native Washington, DC, and by combining it in a new orchestral arrangement...

But it gives no clue how they did it. I thought this was impossible. How might it have been done?

  • If it was only guitar and vox, and they had the original tapes with separate tracks, it's simple. – Tim Nov 5 '19 at 13:13
  • I think if it's like the other Eva Cassidy stuff, then there's no multi-track. There's just a single one-microphone recording of guitar plus vocals. I wouldn't have asked the question if I thought there was a multi-track recording available... – Brian THOMAS Nov 5 '19 at 13:30
  • 1
    You don't need the original tapes but the process would be quite involved. We have (or can make) frequency-time domain profiles of the instruments and use match filters to find them in the mix. Subtract them out and repeat. This may need to be done multiple times on a single instrument and errors can be introduced. It's not easy but possible. Also, if there was only one instrument and a voice and there are blocks of time where they don't overlap this makes it easier. In the end I suspect that the vox would still have other things in it but can be smoothed out with filters. – ggcg Nov 5 '19 at 15:27
  • @ggcg you should make that an answer – Legorhin Nov 5 '19 at 15:40
  • I'll give it a try. – ggcg Nov 5 '19 at 17:48
4

You are asking how it might be done so the following narrative is somewhat speculative but sound.

You don't need the original tapes but the process would be quite involved. We have (or can make) frequency-time domain profiles of the instruments and use match filters to find them in the mix. Subtract them out and repeat.

The model signals could be from a library or you can make them straight from the recording itself if there are segments where the instrument(s) are playing solo. Actually the complement might be an easier process. Focus on the one thing you want, filter it and throw everything else out. The idea is not just to get the notes played by the instrument (if you have a score you can use that too), but to get the acoustic fingerprint of the instrument. One advantage to using the recording to sample different instruments is that you get the effects of the room and everything else included.

This process may need to be done multiple times on a single instrument and errors can be introduced. Some type of adaptive filtering could be employed where the filter parameters are refined with each try. In some cases you don't need a sophisticated model or a perfect sample (which you cannot get anyway).

It's not easy but it is possible. Also, if there was only one instrument and a voice and there are blocks of time where they don't overlap this makes it easier.

It is not likely possible to get a perfectly clean voice or other instrument pulled from a recording. There are other uses of this like narrowing in on a conversation recorded in a crowded place. The separated voice will have some remnants of the other voices present and other sources of noise. But if it's clear enough be coherent various noise reduction techniques can be used to smooth out the voice track. Since it will be used as an overlay to an orchestral recording the left over noise will not be too bad.

As Tim pointed out, if you have isolated tracks this is a very easy operation.

1

Source separation is impossible in the general case but can be approximated for some limited sets of instruments.

If you're interested in the computing methods and currently achieved results, you can check :

https://sisec18.unmix.app/

This website aims to introduce state-of-the-art algorithms in audio source separation. Here we present the Music (MUS) task of the sixth community-based Signal Separation Evaluation Campaign (SiSEC 2018). We show both, objective and subjective results of 50 full-track songs of different styles separated into five components: vocals, accompaniment, drums, bass and other.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.