We listen to the birds from 1905. Recovery of old records

Not so long ago, on Habré, the news that scientists managed to restore one of the very old records, made back in 1905, slipped through without damaging the carrier. The main achievement here is just the fact of the integrity of the carrier, since the recording was made not on anything, but on a wax cylinder. This is almost the very first invented method of recording / reproducing sound, which was widely used. Prior to this, the carrier was glass cylinders with soot (they didn’t know how to reproduce them), then the foil became a coating and only then wax.

Why this is all so cool and technologically advanced, as well as which scientists had difficulty recovering, you can read the original news, and I will tell you about how I became interested in the resulting recording and took it up a little. In general, if you are interested in the process of restoration of any audio-rarity, welcome under cat.

To my surprise, the original recording turned out to be quite noisy, although it was claimed that it was processed by various noise-suppressing algorithms, etc. ... It is worth noting, of course, that compared to other wax-reconstructed recordings, this one really was very good - in quality already quite comparable with the first plates. However, as you know, the best is the enemy of the good.
')
In general, I am a fan of old pre-war songs, and often I have to update the sound of the exhibits from the depths of the network. In particular, for example, recordings of songs from old films suffer from this, since there the original of the film itself does not shine with sound quality.

In our case, we have, in fact, the impression of the track, made by a laser. Now I do not consider the fact that gentlemen, archaeologists also applied some procedures to the original sound, but I will assume that they tried to reproduce the original as accurately as possible. Since the linear dimensions of the track, in this case, are quite large, there is almost no talk of digitization errors, all the more so for the source file a sufficiently high sampling rate was chosen, about four times the frequency of the recorded sounds in the recording. So we can assume that before us is almost a perfect cast of the original sound wave.

Here we are confronted with the peculiarities of the material and the way of reproduction.

Fact number one: the material from which the cylinder is made is quite soft (wax), on it, even if it is perfectly new, it is impossible to capture sounds from a long wavelength less than a certain value.
Fact number two: in addition to the material, the technology of recording on such cylinders makes its own adjustments - the sound was literally written with a needle on the cylinder.
Fact number three: the reproduction of such a cylinder itself destroyed the carrier.

The first fact gives us a maximum recording frequency limit of about 5-6 kHz and, as we will see later, this is very important. And the second and third fact report that it is possible not to worry too much about the preservation of the steepness of the fronts and the shape of the waves - not the accuracy initially.

To begin with, let's turn on the spectral representation of the signal and take up the equalizer (and the equalizer is our everything).



What do we see on the spectrogram? The file sampling rate is already 22000 kHz, while there are no sounds in the recording, as we see, there is no higher than 4.5 kHz, which is to be expected (see the fact number for the times). However, if you take a closer look, some dirt still seeps higher (to better be seen, he increased the contrast and brightness in the square, circled in red). Where it comes from is incomprehensible, but without going into details, in the first place we can, with a clear conscience, pry up everything above the so-called. “Nyquist frequencies” for our sample rate (11 kHz). Since the reserve is still decent there, I didn’t have to waste time on trifles and removed everything above 8 kHz, as well as below 100 Hz, because, judging by the spectrogram, there wasn’t anything sensible either.

After a bit of thinking, I approached this moment even more radically, namely, I did not freeze with the equalizer, but I started everything from the beginning and immediately changed the file sampling rate to 11 kHz.

Next, without further ado, let's use the tool that many modern audio editors have: Noise Reduction. The idea is simple - select a place in the track, where we have nothing but noise. Next, create a pattern of these noises (Capture Profile). In the simplest case, you will only need one single Noise Reduction Level slider.

You can find out about what all the other parameters are for, for example, from this video: An interesting video from Adobe. In English.

True, he says there that they supposedly updated the engine (talking about version CS5.5) and now they know how not to create unnecessary artifacts with noise cancellation, but you and I know that the difference is almost absent. And the settings are the same, except that the window in the new version is redrawn.

There are still some nuances that this charismatic man in the video does not affect, for example, the "window width" for the Fourier transform (FFT Size). If you do not go into details, when getting rid of random (this is important) noise, you should strive to use as many points as possible (Snapshots in profile), either with an equal FFT Size, or approximately two times smaller. All this means that for a high-quality “noise profile” we need as long a segment as possible, in which there is only noise. In general, the good thing about Noise Reduction is that it can be used not only for noise, but also for various background sounds (forest, rain, etc.)

In more detail about all these clever transformations it is written, for example here .

Actually, SoundForge also has more interesting tools, such as the ability to subtract a wave of arbitrary shape from the signal, but I started using Audition, even in bearded times when it was called CoolEdit, and then SoundForge couldn’t even look like , and now I’m too lazy to take on something new.

I habitually took up looking in the file for a segment with noises, but it turned out that they were too short and, when sampling for them, only too short patterns could be made. Because of this, either noise cancellation was not heard at all, or wildest artifacts appeared. Then I began to think how to deal with this. For now I thought, I decided to go from the other end.

The most annoying in our case are the soft clicks, as well as the periodic noise that immediately follows the clicks. I think this effect appeared due to deviations of the cylinder cross-section from the circle shape, or the axis was not exactly standing. At some point during recording, the needle entered the wax too deeply (initial click) and then the short section went out of alignment (the characteristic noise continued several times after the click), then the cylinder made a full turn and the noise repeated. As we will see later, we still manage to get rid of this noise with the help of Noise Reduction, however, moving through the file and looking at the waveform, I noticed that there are also quite strange jambs, similar to the characteristic distortions of the A \ B-amplifiers. A very typical example on the title screen of the article, but there is still quite controversial, since that distortion is too long a period (I chose at random from a file). But the next screen shows very well what I mean.



About how these things ( not ) need to be treated and why, I decided to write separately, and in order for the article not to grow too much, I hid it under the spoiler. You can skip this piece, it is almost a lyrical digression.



In one of the moments when I looked at the spectrogram of the signal, it dawned on me that at the end of the song there is a rather long time when nothing but “whistling birds” sounds. And this whistle on the spectrogram has a completely unambiguous band.



So we can cheat. We uncover the equalizer, set the maximum Range (this is the dynamic range, to put it simply, how much dB will be increased / attenuated by one or another frequency) and cut the frequencies at which the birds sing, and leave everything below / above.

Since even the maximum dynamic range is not enough to completely kill all the birds, I repeat the equalizer twice. In general, about how it works, and why everything is happening this way, you can write a separate article, only, I'm afraid, I don’t know the mathematics of algorithms so well to be clever about this topic.



So, now we have a rather long segment with only noise ... and, correctly, we return to what we started from. We capture the noise profile (after capturing, click close, not cancel, for otherwise all settings will be reset to the previously used ones).

In addition to Noise Reduction, there is also a Hiss Reduction filter, which, as the name implies, will help us get rid of whistles and all sorts of other things. The settings there are almost similar to the noise level, except that the FFT Size somehow works differently, but I didn’t figure out exactly how it is, so I’ve been doing it empirically, and I advise you. For Hiss Reduction, it is also necessary to specify the base noise level (the Get Noise Floor button), and so, this base level should be captured on the same segment where we captured the noise profile.

Next, roll back the last two actions (the equalizer on the birds), then select the entire file and apply Noise and Hiss Reduction to it successively.

After applying these two types of noise, we get a result that is already completely consumable. Is that small artifacts at the ends of the spectrum appear. Here an equalizer comes to the rescue again - we mercilessly cut everything below 150 Hz and above 4.5 kHz.

It became noticeably quieter, but clicks are still audible. Now the spectrometer is running again. If at this stage you listen to the file and watch the spectrogram, it will be obvious that at the time of the click, the noise has a very wide spectrum, but the melody, on the contrary, has clear wavy lines (below, for clarity, I selected one click in red).



To begin with, we will remove the impressive bursts at the places of clicks. To do this, switch to the waveform display mode. As a rule, this is just one “going wild” wave period.



In case this period was just too loud, then I usually just made it quieter. If it was also very distorted, then delete it entirely (too lazy to straighten everyone, which is already there).



And then I will tell you how to literally in two clicks edit the amplitude of such jumps.



After the loudest clicks were defeated, it's time to get rid of the small distortions that exist throughout the file. In the previous spoiler, I already told how to remove them is not worth it, but now I will tell you how to do it more or less correctly.

Here we again come in handy spectral display of the signal. If you look at the toolbar in this mode with a keen eye, a thing that seems to be completely unrelated to the world of audio editors catches your eye. Brush. We need it.

It allows you to remove arbitrary areas on the spectrum. Like a dynamic ultra-accurate equalizer.

Remember I wrote that clicks have a wide range, but the musical sounds are well read on their background? Now we will take advantage of this. Choose this brush, look for the moment where we have such a column of noise on the background of music. Next, draw this brush in places where we only have noise, without affecting the musical line. Next, press del, then move the selection slider to view what we did, or just start painting with a brush in a new place. And we see that where we have just been removed, we now have darkness, that is, silence.



Then we repeat this procedure for all places where clicks are still clearly audible. In principle, with proper persistence, it is possible to uproot everything in general in all places, leaving only the harmonics.

Only this ungrateful occupation, in fact. Because in reality, sounds have a much more complex nature, and in addition to noises between the main harmonics, there are often less important overtones. But despite the fact that they are less important, they give the sound a more natural color and character, which can be lost if you start cutting out all non-harmonic sounds at all.

By the way, remember I wrote about the fact that when periods are removed, there is a certain noise, which seems to be like nowhere to come from? Working with a brush for a long time in the spectrograph, one can notice that even in this mode such noise also sometimes appears - it is not clear where white areas appear around a remote area - this is what it is.

So, as a final touch, we run Automatic Click remover in a mode slightly stronger than the average (upper slider 30, lower 75) - it will remove sharp clicks that could appear as a result of our manipulations. And the equalizer once again cut off all above 5 kHz and below 100 Hz. Next, we normalize the file to 100%. I also deleted the very beginning of the file, literally half a second somewhere, there is still nothing after all our manipulations.

By the way, the original article was not mentioned, and the tags in the file were not affixed, the performers of the original: Harlan And Belmont.

And Byron G. Harlan is even on discogs.com .

- Here, for comparison, the link to the original: yadi.sk/d/sLKZlyp2eQ2YK
And here is my result: yadi.sk/d/lPlJJaEQeQ2Zc

I hope you were interested.

PS I apologize in advance for the lack of axis designations on screenshots. Where the blue - there everywhere vertical Hz, where the green - DB. Horizontally everywhere seconds.

Upd. "Limit on file download exceeded"

Here is this yes! I did not think that Habr even Ya.Disk put.

Here, temporarily laid out on my personal file-trash, hopefully not collapse ...

www.chipinfo.pro/files/Birds.mp3