Singing in the rain

Not me! The birds.

Last year I decided I ought to buy myself an audio recorder, and so I extravagantly did: a Tascam DR-07X, if you’re interested in that sort of thing.

This of course led to downloading the popular audio editor Audacity, and to experimenting with it to see what I could do.

The rain

Here’s the result of such an experiment. A couple of weeks ago we had some nice heavy rain, which was making a nice satisfying noise, so I stood at the back door for a while trying to get a nice satisfying recording of it. What I got was a mixture of rain (wanted), birds singing as though oblivious to the rain (wanted), button presses and wind dropouts (not wanted), and a motorbike (definitely not wanted).

So, here are two versions of the recording.

In the first, all I’ve done is increase the sound level.

In the second, I’ve done my best to remove the extraneous noises while keeping the birdsong intact and not affecting the sound of the rain more than necessary. I think I did better with the first two of these than the third. Have a listen, and see what you think.

I’m particularly pleased with the successful removal of the motorbike.

(The download button in the player will give you a 22 MB CD-quality file of whichever track is currently playing. If you’d prefer something smaller, here are MP3 versions of the original and edited track. They’re 4.1 MB each.)

The editing

The technique I used was as follows:

  • First split the track into two: a high-frequency one containing the birds and the upper end of the rain, and a low-frequency one containing the motorbike and the lower end of the rain.
  • Where the motorbike appears, carefully listen to the rain, and identify a nearby section that sounds similar but has no motorbike.
  • In the lower track only, completely delete the section with the motorbike. Replace it with audio from the motorbikeless section, with a short crossfade at either end.
  • Recombine the two tracks into one.
  • Apply equalisation, to restore the correct frequency balance.

The result is a track in which the birdsong has been left (almost) untouched, the motorbike is gone, but the upper and lower frequencies in the rain are sometimes from entirely different raindrops.

Making the split

To split the sound, I used a high-pass filter on one copy of the original track and a low-pass filter on another, both with the same cutoff frequency. But what frequency?

I could see from a spectrogram view of the original that most of the birdsong was above 2000 Hz, and most of the motorbike sound was below 500 Hz. The sharpness of the filter cut-off is specified in dB per octave, and these pitches are two octaves apart. So I made the split midway, at 1000 Hz: an octave below the birdsong and an octave above the motorbike. This should mean the birdsong and motorbike were equally well removed by the two filters.

(To imagine 1000 Hz, imagine the Greenwich Time Signal pips: that’s their frequency.)

On listening to check, I couldn’t hear any birdsong at all in the lower track and I at least couldn’t identify any motorbike sound in the upper one, as intended.

Would it work?

I was most unsure whether “fake rain” made up of high frequencies from one lot of rain and low frequencies from another would still sound realistic. In particular, would the process affect the stereo spread, built up from the locations of thousands of individual drops? And might the result sound like an unnatural mixture of some muffled drops and some overly bright ones?

In the event, I think it did work, but I can’t quite decide whether the stereo of the original is more convincing.

Equalisation

The equalisation step is needed because the splitting and recombining process is imperfect. Frequencies close to the split end up louder than they should be, resulting in an obviously “boxy” sound. I did this part by ear, and the reconstituted rain definitely doesn’t sound the same as the original rain. However, I reached a point where I preferred the sound to that of the original, so I stopped there rather than try to get a perfect match.

Really I think I should do the maths to work out the theoretically correct equalisation, and use that as a starting point.

On the other hand, the rain still sounds heavy and very wet, which it was, and the birds still sound determined to keep singing, which they were. And I hope you’ll enjoy listening.

Leave a comment