Take the 2-minute tour ×
Sound Design Stack Exchange is a question and answer site for sound engineers, producers, editors, and enthusiasts. It's 100% free, no registration required.

There is a party and as usual people are grouped. We have 2 groups - in first there is a woman who is telling a story at the moment. In second, there is a man, telling interesting anecdote. We have 7 microphones in the room:

  • one is attached to mans head
  • one is attached to womans head
  • five are placed around the room

So - I've got 7 recordings:

  • man.wav (a man telling his anecdote)
  • woman.wav (a woman telling her story)
  • 5 recordings, where voices of two storytellers are mixed

I want to generate a signal from 5 recordings (where voices are mixed), so that mans voice would be louder and more focused then womans.

Can you give me some ideas please? How can I help myself with man.wav and woman.wav? How can these 5 recordings from different angles help me?

Thank you.

share|improve this question
    
Are the recordings synchronized? If so, the "louder" part should be trivial (just mix in more of man.wav than woman.wav). But what does "more focused" mean? –  bzlm Dec 29 '10 at 18:56
    
Recordings are not synchronized. Louder and more focused means just that womans voice must be in the background (or completely eliminated) and mans voice dominant and his speech understandable. –  genesiss Dec 29 '10 at 19:03
    
Ah, so you basically want to use the other recordings to remove the woman's voice from man.wav? –  Matthew Read Dec 29 '10 at 20:07
    
I suspect the answer is going to boil down to using EQ cleverly, but if you can figure out how to filter out a sound source based on a comparison recording, I have a lot of live band tapes with sound source spillover that could benefit from such a technique. –  neilfein Dec 29 '10 at 23:58
add comment

migrated from avp.stackexchange.com Jan 24 at 20:09

This question came from our site for engineers, producers, editors, and enthusiasts spanning the fields of video, and media creation.

5 Answers

The "cocktail party problem" is actually a famous example in signal processing, and there are several different Blind signal separation or Source separation algorithms to solve it in different conditions. Are you doing this for homework in such a class? I've got an example of one on my website, but this statistical method (ICA) wouldn't work well for you because sound doesn't travel instantaneously.

I don't know if there are any more end-user solutions, like DAW plug-ins that do stuff like this, but it's at least theoretically possible.

If this is a one-off, you might be able to do this manually. Try just putting the man and woman on separate tracks and inverting the signal of the woman, then shift the position of the recordings relative to each other (delay of sound is 1 ms per foot) until the two copies of her voice are perfectly in sync (maybe you can see peaks in the waveform display). Then lower the gain of the woman recording until it perfectly cancels out with the part in the man recording. Then you could add all the other recordings and do the same thing with them to make it better.

share|improve this answer
    
+1 for mentioning the cocktail party phenomenon on this site. –  Pelle ten Cate Dec 31 '10 at 15:30
add comment

(I submitted this answer to the SO version of this question, I assume one or the other will be closed, not trying to pad rep or anything)

You might look into the LMS algorithm by Widrow and Hoff. You could apply it to each of your mixed recordings, using man.wav as the desired reference and woman.wav as the disturbance signal. There's also a way to use it for directional antenna arrays, here's an overview.

Also, look into the Filtered-X LMS algorithm. If I recall correctly (which is a big if) it provides better performance in the case of relatively narrowband references, which probably includes voices.

If you don't get to use man.wav and woman.wav, the keywords to look into are blind source separation and higher order statistics.

share|improve this answer
add comment

Why from the 5 recordings? Why not start with the personal microphones of the speakers?

That is going to be way easier. Then you can mix in the other 5 for background noise or raise the volume of these when somebody in the audience is saying something.

share|improve this answer
add comment

If you have precise coordinates of each of the microphones, you may be able to reinforce the mans voice by calculating the time the sound travels to each of the microphones, and applying delays to the various sound tracks.

I would first try combining the signal from the man's microphone and the other 5 microphones

Now do the same for the woman's microphone and see if it improves further if you subtract some of that signal from the other sound tracks.

share|improve this answer
add comment

Not being educated in sound recording I would have to answer this problem in a practical manner and from my real world experience.

If you have a mic on subject A/B and 5 other mics in the room, as a practical means you would want to create a sync sound that you could then take all 7 independent recordings and sync them up in a multi-track editor such as ProTools or Sonar.

I do a lot of video with music shooting and find that you have to have a classic clapper sound or snare drum hit to sync the waveforms. In the video world there is this great plug-in call Plural-Eyes that do the leg-work.

It's not so hard once you sync up all seven recordings in ProTools to mix them correctly.

Now, if this is a theoretical exercise, smarter people than I have supplied answers. ;-)

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.