This is good feedback. I've become accustomed to listening to heavily processed audio for voice communication. So maybe I find it easier to listen to the processed audio? It was fun and a small learning experience for me to do this either way.
I should have said that I thoroughly enjoyed the writeup by the way. The end result has a tinge of the kind of artefacts that you get with stem separation to separate vocals from music, I might try putting the original through Brusfri later to see how it does - https://klevgrand.com/products/brusfri
Thanks. Part of the goal was to apply some signal processing stuff I learned. The other hope was someone who knows more about audio preservation than I do would get access to an extant B-58 and see if they could produce better quality recordings of this. It is a unique part of history we take for granted today I think.
No, I agree. There’s an impressive level of theory on display here, and I admire the passion, analysis, creativity, and follow-through. The end product matters less than the journey, but I agree with you that it’s missing the discernment of a human ear practiced in this kind of work.
Maximizing for signal-to-noise ratio sacrifices the integrity of the voice, with that telltale “underwater gurgle” quality to what’s left of it.
The same thing happens in photography: noise removal software removes the static but also causes the photo to lose detail. The human brain is exceptionally skilled at seeing and hearing through noise.
It could have been recorded in the aircraft (running under its own power or on shore power), or in the shop running off a bench AC power supply. Either way, if the unit's power supply filter capacitors are degraded (which is very likely), then the AC power supply frequency will leak into the audio output. I deal with this nonstop servicing audio amplifiers from the same era as the the B-58.
It sounds a lot like a whine from an old motor, maybe the one spinning the tape reels. My guess is someone hacked the output to a speaker and stuck a microphone next to it to record, and the hiss/whine comes from the tape machine itself.
> Several sections appear in red, this means they are "0 dB" which indicates a maximum audio level. This is why the recording is distorted
I think it's because the signal is normalized, there's no aliasing in the clip which would be the telltale sign of digital clipping.
> There is to my knowledge no general algorithm to remove noise from an audio sample.
There's quite a lot of research into denoising and blind source separation, simple filtering is kind of a start but as you can tell from the demos doesn't necessarily make it better. Search those terms in the IEEE transactions on DSP/Audio and jAES.
> I also applied a high pass filter passing all signals above 260 Hz then a low pass filter passing all signals below 2850 Hz. This reduces the audio signal outside the range of normal human voice, which is probably just noise anyways.
This is true in this clip but this description is kind of bad. The human voice absolutely has energy about 2.6k (in fact that's right around where our ears are most sensitive) but this clip does not. Old commercial audio systems had very bad bandwidth due to the limitations of the wiring/transmission, so microphones that typically plugged into them often had what we'd call "bad" frequency responses with higher bass rolloffs (to avoid distortion mangling the voice), lower treble rolloffs (because the materials were cheaper and they didn't care about frequencies that couldn't be transmitted).
Basically, if you know what's going to play back the voice is going to be distorted and have limited bandwidth you intentionally highpass filter out the fundamental and accentuate the middle range of the voice which is critical for intelligibility. Paradoxically distortion can help with intelligibility in low doses.
This kind of stuff is why restoration is hard, because if you're trying to recover what it sounded like in the room you need to consider the recording system as much as the characteristics of the recorded noise. And it's impossible to know, so some guesswork is needed.
In case you're not aware, NASA released several thousand hours of Apollo 11 audio which is similarly noisy. You might have a good time messing with some of those recordings.
Someone needs to add a sound sample to the voice warning system article[0] ASAP. Also, Gina Drazin isn’t mention there, but should be with that ad as supporting evidence.
Reminds me of when we got a local TV/radio newsreader (Julie Skentelbery) to voice the (default) TERPROM Terrain Awareness Warning System (TAWS) alerts (which no-one ever hears because they all have their own preferences):
I hadn't poked around with Adobe's latest AI audio podcast enhancement recently so I was curious how it would compare and was pretty impressed. Results here - make sure to tap the unmute icon.
34 comments
[ 2.6 ms ] story [ 83.3 ms ] threadThat said, I love a good applied math write up. That we can also evaluate the result using human perception just helps keep us honest
Maximizing for signal-to-noise ratio sacrifices the integrity of the voice, with that telltale “underwater gurgle” quality to what’s left of it.
Just so much historical charm, hard to put into words.
* https://fcxinc.com/why-the-aviation-industry-operates-on-400...
I wonder if when these recordings were made someone was actually powering up the whole aircraft by running an external APU of some kind?
https://en.m.wikipedia.org/wiki/Wiener_filter
I think it's because the signal is normalized, there's no aliasing in the clip which would be the telltale sign of digital clipping.
> There is to my knowledge no general algorithm to remove noise from an audio sample.
There's quite a lot of research into denoising and blind source separation, simple filtering is kind of a start but as you can tell from the demos doesn't necessarily make it better. Search those terms in the IEEE transactions on DSP/Audio and jAES.
> I also applied a high pass filter passing all signals above 260 Hz then a low pass filter passing all signals below 2850 Hz. This reduces the audio signal outside the range of normal human voice, which is probably just noise anyways.
This is true in this clip but this description is kind of bad. The human voice absolutely has energy about 2.6k (in fact that's right around where our ears are most sensitive) but this clip does not. Old commercial audio systems had very bad bandwidth due to the limitations of the wiring/transmission, so microphones that typically plugged into them often had what we'd call "bad" frequency responses with higher bass rolloffs (to avoid distortion mangling the voice), lower treble rolloffs (because the materials were cheaper and they didn't care about frequencies that couldn't be transmitted).
Basically, if you know what's going to play back the voice is going to be distorted and have limited bandwidth you intentionally highpass filter out the fundamental and accentuate the middle range of the voice which is critical for intelligibility. Paradoxically distortion can help with intelligibility in low doses.
This kind of stuff is why restoration is hard, because if you're trying to recover what it sounded like in the room you need to consider the recording system as much as the characteristics of the recorded noise. And it's impossible to know, so some guesswork is needed.
In case you're not aware, NASA released several thousand hours of Apollo 11 audio which is similarly noisy. You might have a good time messing with some of those recordings.
https://apolloinrealtime.org/11/
https://en.m.wikipedia.org/wiki/Voice_warning_system
https://www.bbc.co.uk/programmes/p0h6dn7x
<https://www.collinsaerospace.com/what-we-do/industries/milit...>
https://github.com/pgerhardt/B58/blob/main/README.md