Is it possible to restore audio that has been tonally modified by a phase vocoder?
-
- KVRer
- Topic Starter
- 3 posts since 25 Sep, 2019
My objective is to apply some process to the modified audio that allows me to obtain the original audio. I think it can be solved through machine learning techniques. I don't know if there is any method or algorithm that solves it. I appreciate any information, pappers, etc. Regards
-
- KVRAF
- 2087 posts since 24 Jun, 2006 from London, England
Likewise is there a way of doing the reverse of a ring mod. In Max I know you can simply multiply the carrier and a sine wave to create a simple ringmod, so assumed 'dividing' by a sine wave would 'undo' it but didn't have much luck!
-
- KVRian
- 626 posts since 30 Aug, 2012
The title of this thread and your text seem to be asking two different things. Are you, A) trying to restore the original waveform - or B) trying to restore the TIMBRE (tonality) of the original waveform?gabrielg wrote: ↑Wed Sep 25, 2019 3:04 pm My objective is to apply some process to the modified audio that allows me to obtain the original audio. I think it can be solved through machine learning techniques. I don't know if there is any method or algorithm that solves it. I appreciate any information, pappers, etc. Regards
A) If you know - or are somehow able to detect - the amount of pitch and formant shift that had been applied you could apply the inverse with another vocoder. However, AFAIK, phase vocoders always distort the signal in order to "bend the laws of physics". Running the signal through another phase vocoder may restore the original pitch but would likely add even more tonal degradation. In other words I believe the process is irreversible/"lossy".
B) There are some processors out there that claim to remove mp3 artifacts such as UnChirp (http://www.zynaptiq.com/unchirp/ that may also help restore the timbre of pitch shifted signals if that's what you're after. Haven't tried it.
-
- KVRist
- 81 posts since 24 Oct, 2000 from Bremen, Germany
Undoing ring mod by dividing by the sine wave the signal was multiplied by could work, but it has to be exactly the same frequency and phase of sine wave - that could maybe be worked out by finding regular zero crossings in the signal - and when the sine wave is zero there is no way to know what the input signal level was, so you'd have to interpolate from the previous and following samples to fill in the gap.
Undoing phase vocoding sounds hard - even if it's the simplest possible phase vocoder (no phase resetting on transients) the output is a mix of overlapping analysis frames so you can 't cleanly get the output of each frame to undo what was done to it.
Undoing phase vocoding sounds hard - even if it's the simplest possible phase vocoder (no phase resetting on transients) the output is a mix of overlapping analysis frames so you can 't cleanly get the output of each frame to undo what was done to it.
- KVRAF
- 15274 posts since 8 Mar, 2005 from Utrecht, Holland
This task sounds like reconstructing a living cow from a pile of incinerated hamburgers.
We are the KVR collective. Resistance is futile. You will be assimilated.
My MusicCalc is served over https!!
My MusicCalc is served over https!!
-
- KVRer
- Topic Starter
- 3 posts since 25 Sep, 2019
AUTO-ADMIN: Non-MP3, WAV, OGG, SoundCloud, YouTube, Vimeo, Twitter and Facebook links in this post have been protected automatically. Once the member reaches 5 posts the links will function as normal.
I regret not having been clear enough, my goal is more like option B). I am not interested in recovering exactly the same waveform. I would like to obtain the same behavior in terms of the frequency of the original signal, for example; if I wanted to identify a speaker, it is crucial that the fundamental frequencies of vocal tract resonance are correctly located in the spectrumFender19 wrote: ↑Thu Sep 26, 2019 11:10 pmThe title of this thread and your text seem to be asking two different things. Are you, A) trying to restore the original waveform - or B) trying to restore the TIMBRE (tonality) of the original waveform?gabrielg wrote: ↑Wed Sep 25, 2019 3:04 pm My objective is to apply some process to the modified audio that allows me to obtain the original audio. I think it can be solved through machine learning techniques. I don't know if there is any method or algorithm that solves it. I appreciate any information, pappers, etc. Regards
A) If you know - or are somehow able to detect - the amount of pitch and formant shift that had been applied you could apply the inverse with another vocoder. However, AFAIK, phase vocoders always distort the signal in order to "bend the laws of physics". Running the signal through another phase vocoder may restore the original pitch but would likely add even more tonal degradation. In other words I believe the process is irreversible/"lossy".
B) There are some processors out there that claim to remove mp3 artifacts such as UnChirp (http://www.zynaptiq.com/unchirp/ (http://www.zynaptiq.com/unchirp/) that may also help restore the timbre of pitch shifted signals if that's what you're after. Haven't tried it.
- KVRAF
- 8828 posts since 6 Jan, 2017 from Outer Space
Is this for forensics? Its a bit like that guy who used a well known process to distort his face into a swirl which was easy to reverse. If you have an idea what kind of process it could have been, you can do some trial and error to reverse it. You need to get into the mind of that guy, those are usually simple minded and possibly used a simple method...
-
- KVRer
- Topic Starter
- 3 posts since 25 Sep, 2019
hahaha I'm afraid it's impossible. Starting only from the output of a system, it is very difficult to establish the transfer of it ... that is why the first solution that occurred to me was the use of a machine learning algorithm that can be trained to reconstruct the signal, which that is, use a phase vocoder in reverse (undo what the first phase vocoder did) to obtain a result as close as possible to the original signal
- KVRAF
- 15274 posts since 8 Mar, 2005 from Utrecht, Holland
Yeah, like the hip answer to any difficult computational task is.... Machine Learning!
We are the KVR collective. Resistance is futile. You will be assimilated.
My MusicCalc is served over https!!
My MusicCalc is served over https!!