BACKGROUND
I have two audio recordings of a meeting. Each recording was taken by a different handheld recorder.

PROBLEM
The recorders each have an integrated mic so there is a lot of noise from handling the recorders.

OBJECTIVE
I want to get rid of the handling noise and end up with a single clean recording of the meeting.

----

SOLUTION ?
As the handling noises will be unique to each recorder then I figure that what I need to do is extract and save the frequences which are coomon to both recordings.

How can I do this? Is there some software which can do it?

Thanks for any info.

Moonshot

Sounds like only natural intelligence, i.e. a wetware brain could do this

Well, the first of the difficulties would be exact synchronisation of the two recorders. Maybe some kind of vocoder can filter one recording with the spectral envelope of the other one, or two vocoders in series fed with both signals. But sure there is no one-click-software that can do this

You do realize that the recordings are not in sync, right? Their speeds (like any tape recorders) are slowly drifting over time.

I'd say do a spectral noise subtraction on each one and choose the one that sounds better.

You say the recordings are not in sync. But is the situation quite as hopeless as you seem to suggest. Perhaps I gave the wrong impression that I had used tape. What I have are two recordings each made on a separate digital recorder. (See below for exact details of the recorders.)

With all the digital cleverness in these sorts of device I guess there would be little drift over time. Although there may be other problems. For example, compression may not have been applied by each recorder in a similar way to low and high amplitude sounds.

The handling noise of the recordings is awful and anything which reduces it would be a great help. However I have almost no hands-on experience with denoisers and am not clear what you mean by "spectral noise subtraction".

Maybe I can set a denoiser to remove frequencies which do not match speech characteristics. Could someone kindly recommend a suitable denoiser for me to try in these circumstances. The discussions here about denoising seem to deal with repeating noise such as hum and not with handling noise.

Thank you.
Moonshot.


==========================================
My two recorders are:
(1) A basic MP3 player using a SigmaTel STMP35XX chip. This is a heavily integrated chip containing almost all the essential functions of an MP3 player. It records voice in either WAV or MP3 format. More info if you need it is at: http://www.sigmatel.com/products/portable/audio/stmp3500.asp.
(2) An Olympus voice recorder which records to digital memory. It records voice to WMA. More info at: http://www.olympusamerica.com/etc specification

An interesting problem. I think my solution, assuming handling noise is relatively infrequent, would involve finding characteristics of the handling noise. Level-match and time-synch both recordings, then scan one recordings for signs of noise. Wherever noise is found, copy the affected area from the other recording.

I don't know how you would automate this in off-the-shelf software or which program would be suitable. However, it should be easy and not too time-consuming to do by hand in any audio editor that allows you to have two files open at the same time.

To me this sounds like a (rather) simple editing job. Transfer both recordings to a pc. Use an editing application to align them (they won't be perfectly sync, but that doesn't really matter). Use one of the two as a "master" track and replace all handling noises by the sound of the other take if it's quieter on those spots.
This requires some basic editing skills, but it shouldn't be difficult nor very time consuming.
As an alternative you can use a sound restoration application. There are a few good but expensive ones. You might be able to find cheap or even free ones that can remove the noises to some extent.
Good luck.

Multitrack wave editors are suited for this job, like Samplitude and Wavelab. I wonder if recent Cool Edit and Soundforge version feature this. Pro Tools could certainly do this too, but I don't know if it is usable as a standalone wav editor.

In these editors, load the two files in a two-tracks project, stretch them so that they are in synch (very easy with Samplitude : just drag the end of the longest one where the shortest one ends ). Then get the drawing volume tool, and draw a level drop everytime one recording makes noise, while drawing a similar level rise on the other recording in order to keep a constant overall level.

I'm sorry Pio, I probably didn't express myself clearly enough. I meant a tool to automate it, without much or any user input required, in case this is a common problem and he expects to have several hours worth of noisy recordings to clean up eventually. All the programs you suggested are certainly good choices for doing it by hand though.

Actually, after some thought, the OP's idea of extracting common frequencies might not be so bad either, although it requires more computational work. Do FFTs of both files over small intervals, decrease the coefficients of frequency groups where the variations are large between recordings, then perform IFT.

I'd like to have a go at implementing something like this, or perhaps both methods to see which works best, but I'm swamped in work at the moment. Perhaps later this week or next week. Moonshot, would it be possible for me to get copies or extracts of the two files for testing?

[EDIT] Come to think of it, even just averaging the two recordings should cut handling noise in half ;-)

Hi fellas, I am the OP.

I have been away for a few days and see that there has been a useful discussion in my absence that might actually point to a workable solution. I was getting worried that I had made two recordings and I was not going to get better quality than one recording!

I have a new suggestion to put forward which may be old-tech but might help do the job. I am nothing like as experienced in audio as you guys, so I can't tell.

The idea is to combine the signals using some ANALOGUE method. Perhaps using some reverse phase. I admit that I am a bit vague on the actual practicality but I hope you can see what I mean. Is this a workable idea as I guess that simple reverse phase would cancel out the very signal I want!

if you used analogue then the audio parameters during the playback of the two combined signals could be varied using manual controls such as letting one or another recording dominate, letting one be relatively louder, etc. The final cut's optimal settings would be determined by nothing more scientific than the operator's ear determining at what point the clarity is best. Of course, the twiddling of knobs would have to be in real time as the audio played.

On the other hand, maybe I all far too old-fashioned. :-) I don't know if modern digital tools are sampling so fast and working so cleverly that the digital tools can easily mimic the analogue combining of signals. And maybe the digital approach permits rather more additional signal manipulation too.

Well. What do you think of the analogue idea when compared to the digital solutions being discussed here?

Best wishes and thank you: you are helping me a lot already, especially as the recordings are very valuable to me. I'm going off now to take a look at the applications Pio mentioned such as Samplitude, Wavelab, and Sound Forge to see if I can handle them!


Moonshot.

I was thinking of something along those lines: synchronising and then adding the two tracks' samples; one would think initially that the actual voice signal would get louder. But the original noises from both players would still be there and it may in fact make it worse in the end. Something like your first idea is probably more likely. Surely there are programs for similar purposes out there already?

Anyway, I don't think the cut-and-paste method would be very effective as the transitions could be very noticeable, the two recordings could sound too different, or other problems could occur.

If the two recordings are perfectly sychronised and added together, the uncorrelated signal (noise) will not increase in level.
The correlated signal (wanted) will add vectorially and increase in level by 3dB.
This is equivalent to a 3dB increase in signal to noise ratio (2:1).
This could be achieved with any software that has stereo to mono mixing.
Since the clocks of the two digital recorders were not synchronised this may not be possible?

Hey, what's the matter with me!

I clean overlooked the sentence where you, Gristus, asked about getting a copy of my audio files to try out how they can be combined. Now that really is the best offer I have had all week. Well, all month actually! :-) Especially as denoising the files is very useful to me.

Yes, I would be happy to send you the files. I hope we can keep them confidential.

• One is a stereo WMA (ugh) but both channels are effectively identical and the format settings permit a lot of information for what is no more than a speech recording: 44,100 kHz at 64kbps.
• The other is an MP3 at 16,000Hz and 16kbps which I would like to think is good enough for speech.

In fact the recorders were used covertly so there is a degree of muffling as well as the handling noise!

The 'meeting' ran for about 30 minutes so I will need to extract some samples.

Do you want me to post them in the Uploads forum here? I guess that means each sample must be only 30 seconds. Is that OK or do you want a longer pair of samples which I guess I would have to email you?

From your point of view, presumably it is best if the samples are encoded back into the compression format they are currently in now. Is that correct?


Moonshot

Don't worry, of course I won't share the information with anyone. A 30 second extract or two should be enough, try to find segments with handling noise representative of the whole recording. The format of the files is not important, I'll convert them to WAV once I get them anyway as it's easier to work with. I'll send you a PM with my email address.

I'm still quite busy with work, so I probably won't have time to look at it until next week though.