Hacker News new | past | comments | ask | show | jobs | submit login
Audio AI: isolating vocals from stereo music using Convolutional Neural Networks (towardsdatascience.com)
275 points by turbohz 5 months ago | hide | past | web | favorite | 71 comments



This is an awesome project, but it seems it was done without reference to academic literature on source separation. In fact, people have been doing audio source separation for years with neural networks.

For instance, Eric Humphrey at Spotify Music Understanding Group describes using a U-Net architecture here: https://medium.com/this-week-in-machine-learning-ai/separati... - paper at http://openaccess.city.ac.uk/19289/1/7bb8d1600fba70dd7940877...

They compare their performance to the widely-cited state of the art Chimera model (Luo 2017): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5791533/#R24 with examples at http://naplab.ee.columbia.edu/ivs.html - from the examples, there's significantly less distortion than OP.

Not to discourage OP from doing first-principles research at all! But it's often useful to engage with the larger community and know what's succeeded and failed in the past. This is a problem domain where progress could change the entire creative landscape around derivative works ("mashups" and the like), and interested researchers could do well to look towards collaboration rather than reinventing each others' wheels.

EDIT: The SANE conference has talks by Humphrey and many others available online: https://www.youtube.com/channel/UCsdxfneC1EdPorDUq9_XUJA/vid...


People have also been doing audio source separation effectively for years without neural networks.


It's interesting cause I have a recording of human voices plus a background TV show that was too loud; I've looked around for something that would be able to separate the two but I haven't found a straightforward solution.

For example if you Google then FASST is one of the ones that come up, but it's a whole framework and in order to use it you'd have to learn the research yourself; much of these software is not geared for end users.



Learn how to do waveform inversions - if you have a stereo signal, anything not fully-centered will come through better while the rest is cut out. You can then take that, invert it, and play it back with the original, cutting out that noise and keeping the fully-centered things like vocals present.

This is how I play guitar to my favorite songs on my computer.


this sounds interesting. could you elaborate a bit? it is unclear if you are inverting once, or twice. "if you have a stereo signal, anything not fully-centered will come through better while the rest is cut out" -- is this before or after an inversion?


Of stereo tracks L and R, you invert R and add it to L, effectively canceling anything centered. This usually removes voices. If you subtract (invert then add) the result from the original L and R tracks you get centered sounds only. Results range from perfect to not effective at all depending on the songs you apply it to.


> but it seems it was done without reference to academic literature on source separation

Sums up towards data science pretty well.


What motivates people to invent phrases like "perceptual binarization" when googling "audio binary mask" literally gives you citations in the field that have been doing this for years?

For example, 2009 musical sound separation based on binary time frequency masking.

Or more recent stuff using deep learning. Also the field generally prefers ratio masks because they lead to better sounding output.


I know from my own experience that it's possible to dig yourself quite deep into some niche research field without realizing that there's an existing body of knowledge about it. If you or nobody else in your research circle knows the right keywords to enter into search fields it's really easy to overlook piles of published papers.

I want to say things were different back when we relied more on human librarians in searching for literature, but unfortunately history is full of cases where people independently discovered the same things as well.


One extreme amusing/alarming example of this: “A Mathematical Model for the Determination of Total Area Under Glucose Tolerance and Other Metabolic Curves”

https://academia.stackexchange.com/questions/9602/rediscover...


My approach to avoid this is always to try and find a recent and well written Masters or Ph.D thesis in the area. You can't always find them of course, but if you do they tend to have pretty good context and a more detailed bibliography than you'll find elsewhere.

That said, if you are still at the point of inventing new terms for things people have been doing for decades, you are probably being fairly superficial in the area as well.

Research areas like CNN are especially prone to this because it is so much easier to apply the techniques than to understand the problem domain, and it generates a lot of low quality research papers. See also "when all you have is a hammer".


> My approach to avoid this is always to try and find a recent and well written Masters or Ph.D thesis in the area.

How do you find this?


Typically start by looking at lab/group home pages who are active and strong in the area . Some universities have good seach capabilities and digital copies, others don’t. Lots of people will have draft versions on their pages though.

If you find someone very active in the area and you like the look of their papers, you can always try asking them...


Early on building https://faunadb.com/, it took us about a year to discover that the literature referred to "historical" or "time travel" queries as temporality. Other startups in our space were also using homemade jargon.

We figured somebody else had for sure done this and kept searching for different keywords until we picked up a bunch of papers with the correct terminology.


Also, the time investment to learn a congruent field's jargon is often much greater than just making up words and let the internet peanut gallery sort out the synonyms for you.


Maybe this is a Cunningham’s Law type attempt at finding prior art for a new patent.


If you look smart and inventive you appear higher social status. Helps nerds impress women I reckon.


Hello, a little self promotion, you can see it our experiment with some deep neural networks doing real-time audio processing in the browser, using tensorflow.js

http://gistnoesis.github.io/

If you want to see how it's done it's shared source : https://github.com/GistNoesis/Wisteria/

Thanks


Does anyone know if this is related to the new iZotope RX 7 vocal isolation & stemming tools? It does seem to be talking about something similar, especially when it mentions using the same technique to split a song into instrument stems.

(Or to put it another way - there is commercial music software released in the last year that lets you do this yourself now.)

https://www.youtube.com/watch?v=kEauVQv2Quc

https://www.izotope.com/en/products/repair-and-edit/rx/music...


Going back further, X-Tracks did this ~5 years ago https://vimeo.com/107971872

That said, I think a deep learning approach will likely do a lot better (and be a lot easier to develop, imo)

Also, check out Google's Magenta project; it aims to use ML in various music / creativity projects.

I personally plan on doing a project that will involve audio source separation as well as sample classification; a good trick for analyzing audio data is to convert it into images (maybe with some additional pre-transformations applied, such as passing it through an audio filter that exaggerates human-perceived properties of sound) and then just use your run-of-the-mill, bog-standard, state-of-the-art image classifiers on the resulting audio spectrogram with some well-chosen training/validation sets.


That's interesting, are you going to make something like LabelImg[1]? I've been looking for something like that for audio, yet I'm not sure about treating audio as images. I've heard of this trick, but NN for audio better do work with RNN, GRU[2], maybe LSTM; and images are processed with CNN.

[1] https://github.com/tzutalin/labelImg [2] https://en.wikipedia.org/wiki/Gated_recurrent_unit


I was gonna do something involving about 3 different neural nets:

a source separator: taking one audio stream as input and producing a set of audio streams as output.

a segmentation regression neural net: takes an audio stream as input and returns start and stop timestamps of individual samples as output, or alternativey, just trimmed copies of the audio stream

sample classifier: takes an audio stream and then returns “kick drum”, “snare drum”, “voice”, “guitar”, etc

then the pipeline would be like

source separator => segmenter => sample classifier

Hopefully with this I would be able to decompose music into constituent parts, useful for remixing and other kinds of musique concrete

I expect that the results with a deep pretrained generic image model + some tweaking with more niche training examples will be satisfactory, but if not it would be a good excuse to experiment with more traditionally seuqnece-oriented network architectures


Do you have to convert it into an image? What is it about classifiers that require image input? I've always found it very cool that audio compression and image compression end up using similar frequency-space techniques sometimes.


You don't have to. It's just that there are a lot of good image classification network architectures that exist already so you can more or less grab a good network off-the-shelf, then give it more targeted training examples so that it performs well in your specific use-case.

I guess it comes down to the fact that a lot of these signal processing techniques involve convolution.


I used to work in an audio processing research center back in 2003, and colleagues next to me were able to isolate each instrument in a stereo mix live using the fact that they were "placed" on different spot in the stereo plane.

Don't ask me how they did that, it was close to magic to me at that time, but i'm sure it wasn't neural networks. Although it probably involved convolution, as it is the main tool for producing audio filters.

If anyone has more info on the fundamental differences of the neural network approach compared to the "traditional" one, i'd be thankful.


Back when I was a teen, I used to strip out the vocal tracks on stereo music, by disconnecting the speaker ground wires from the back of the amplifier and connecting them to each other (so each speaker would ground against the other one). Since the vocal was "center", it had the same waveform on both speakers, so the speakers couldn't make a sound if they couldn't dump the ground to something. And the instrumentals weren't distorted too much. At least it worked good enough for a cheap Karaoke setup.


I discovered this in the 80s when wires in to my car stereo speakers came loose and suddenly the vocal or lead instrument was missing from the music. Really puzzled me for a bit til I found the problem.


You can generally do this pretty closely digitally by taking the left and right channels of a stereo track and inverting the phase on one of them. Since the vocals are usually panned center (as you said), the inverted phase ends up destructively cancelling them. Depending on how the instrumentation was mixed, usually the instruments are left pretty in-tact.


Usually the track sounds awful. For example, the bass can go away with the vocals, instruments' timbre change etc.


Makes sense since panning bass would be an odd thing to do.


I discovered something similar by accident, only with the line audio cables between by tape/CD and the amplifier. I would pull the plugs out just enough so the tip made connection but not the ring. It would give the same effect.


There's a trick you can do to isolate vocals from some music by essentially flipping one of the stereo channels and combining their waveforms. All the stereo data cancels out and you're left with anything not panned hard center. Recombine that with the original stereo file converted to mono and you then get the vocals, usually a bunch of cruft from the reverb, and anything else panned hard center


Most smartphones shift phases on the left/right channels to make it sound better in headphones (simulating bigger room), this means you can combine the two channels (by connecting the positive wires or just wiggle a little in the headphone jack half way in) to extract the vocal track. Not as fancy as a CNN but I’m guessing it would be better to do something with the stereo information in preprocessing before training.


Removing the center is possible, but how exactly would you "recombine" it with the original to keep the center? Correct me if I'm wrong, but I don't think the math works out like that.

Say we have mixed three sources, let's call them L (panned hard left), C (center) and R (hard right). Then the left channel has +L+C, and the right channel has +R+C.

Now we phase invert one of them, say the right channel, and combine them. The new mono file is +L+C-R-C. +C-C cancels out and we're left with +L-R.

Since +R and -R essentially sounds the same, it sounds like if we had originally done a mono mix of L and R (+L+R).

But we can't combine this with a straight mono conversion (+L+C+R+C) in any way that will remove both L and R. All we can do is reproduce +L+C or +R+C.


A stereo audio file only has two channels, mixed from any number of sources: L+R. This described process essentially converts the track (we'll call X) to mid/side (M/S) form, which is used by some microphones and other outboard equipment. We're exploiting properties of both forms of the audio to produce our isolated center.

Also if you have the removed data all you have to do is invert it and sum it against hte original material and it will cancel out

I described it slightly incorrectly in the original post but I have used this process to extract vocals for remixing

The first inversion produces one channel containing nothing but stereo data, as you're summing L+Ri to produce S -- the center channel data (technically, anything equivalent in both stereo channels) cancelled itself out. (Note that if you combine L+Ri you should get a zero-ed out waveform)

The second inversion, produced by combining the first inversion (Ri) and the original source converted to mono (L+R), produces the isolated center channel (M). It works because you are essentially only cancelling out stereo data you generated in the first inversion

If you don't believe me, load up a copy of Audacity or Sound Forge and try it. (Note the music needs to be uncompressed) One track this works with is the original mix of "Day n Nite" by "Kid Cudi" if you can find a WAV copy somewhere. It doesn't work with a lot of music

So...

  L, R = stereo_split(X)

  Ri = inverse(R)

  S = L+Ri

  Si = inverse(S)

  M = Si+(L+R)

  Note: L+Ri should result in empty waveform


Trying to follow this, but I’m lost at a few points.

If you say you have a file where +L-R = 0, then L = R? That’s just a mono file?

L/R stereo and M/S stereo conversion works like this. M = L+R, S = L-R (or S = R-L, either one is fine).

If you expand out your expression M=Si+L+R, where Si=-(L-R)=-L+R, you have M=-L+R+L+R, which means you are saying M=+R+R. This doesn’t seem right.

If you just want an actual M, that’s just a mono conversion of L+R. For sure, if the recording is M/S, the two sides cancel out and you will extract the mid channel, which would often contain song - but there’s no reason to do all the repeated inversion and mixing to do it!


+L-R = 0

+L+R gets you a mono file (before volume reduction). +L-R cancels out. Try it out in a wave editor with a sine wave

The figures I wrote in my previous post use + as a summing operation, not to signal polarity

I discovered the process on my own, so I'm sure there are better ways to do it, I'm not a recording engineer :)


Interestingly, with a Nexus 5 and a certain pair of earbuds, this seems to happen automatically (minus the last step) - with most songs, I can hear the backing tracks, but no vocals. The earbuds are passive components, so I assume there's something going on electrically that's combining the stereo channels somehow.


This works especially well for pop tunes, where the vocal track is by far the most prominent and is almost universally panned dead-center in the mix.


Almost certainly something like "Real-time Sound Source Separation: Azimuth Discrimination and Resynthesis" (Barry et al, https://arrow.dit.ie/argcon/35/)

This is a more sophisticated generalisation of the idea of inverting one channel and averaging in order to isolate the centre-panned vocal (or remove it for karaoke purposes).

It works well with mixes in which stereo placement is entirely by pan pot, adjusting the left/right volume levels of each instrument individually in order to place it on the stereo image. It doesn't work so well with real room recordings, where stereo placement is determined also by timing information (read e.g. https://www.audiocheck.net/audiotests_stereophonicsound.php)

This is a better-specified problem than monoaural source separation, which I think is what the original article here is about.


I'm a fairly n00b-level hobbyist music producer, but I can take a stab.

I like thinking of the "stereo image" of a track as having 3 dimensions in physical space. If you imagine looking forward at a stereo system, left to right is the "pan" (how much of a signal is originating from the left vs. right speakers, with 50% each being dead center), volume of each track (or instrument) is how distant (or close) the sound is (you can imagine an individual instrument moving forwards toward you if louder, and vice versa), and the frequency (or pitch) is how high the track is (with lower pitch or frequency being at the floor).

When a mixing engineer "mixes" a track, each track (or instrument) tends to be: 1) adjusted to an appropriate volume relative to other tracks (forward/backward), panned or spread left/right (usually more stereo "width" on the higher frequency sounds), and "equalized" to narrow and manipulate the band of frequencies coming through to the master mix from that track (it's common to cut out some of the harmonics either side of the fundamental of one instrument to leave "space" for another instrument competing for the same frequency range).

So now, even with a final, mastered track, we could apply various filters to de-mix fairly easily if the mixing engineer has left a good amount of "space" in between the elements in this stereo image. If our bass is always under 150 Hz and every other element is above 200 Hz, a simple low pass will grab the bass (and likely also the kick drum, if there is one). But we could also use a "gate" to only allow a signal of a certain magnitude (volume) through to isolate that kick hit, or the inverse to exclude the kick and just get the bass sound. A band-pass could do a decent job on a guitar or vocal track, but will also grab background noise from other tracks sharing the same frequency range. More complicated techniques could be used to isolate things that have been panned left or right of center based on comparison between the left and right stereo channels of the mix.

These techniques won't be as universally useful as the approach in this article, but for certain tracks or sections of tracks, can give very good results very quickly. More tracks and effects like reverb and distortion make all of this more difficult to do with simpler techniques.


I think you’re overestimating how independent each instruments section of the frequency spectrum is, even after mixing engineers cutting EQ to make places like you said.

The reality is, even instruments like kick drum or bass guitar have significant tonal content above 200hz, much of it overlapping with the guitar and vocal ranges.


>using the fact that they were "placed" on different spot in the stereo plane.

From the description it looks like they used FastICA. https://en.wikipedia.org/wiki/FastICA

The traditional is faster and less recourse intensive. This approach is just showing that it can be done at this point.


Historically, one approach has been independent component analysis (https://en.wikipedia.org/wiki/Independent_component_analysis).


you can compare the L and R channels to each other to separate signals out of a stereo field. https://cycling74.com/forums/separating-stereo-segments-poss...


Trivia: Avery Wang, the guy who invented the Shazam algorithm and was their CTO did his PhD thesis on this topic:

https://ccrma.stanford.edu/~jos/EE201/More_Recent_PhD_EEs_CC...

``Instantaneous and Frequency-Warped Signal Processing Techniques for Auditory Source Separation'' (1994)


Please stop publishing on Medium. I'm getting error "You read a lot. We like that. You’ve reached the end of your free member preview for this month. Become a member now for $5/month to read this story".

Not gonna do that.


Share this sparingly lest it gets "fixed" ;) https://outline.com/SauXTY


There is a lagged autoregressive technique used in forensic analysis that allows 3d reconstruction using 1d (mic) sound.

A CNN should be able to back that out too, and do other things like regenerate a 3d space. In the right, high-fidelity, acoustic tracks could be the spatial information to reconstruct a stage and a performance. It would be neat/beautiful/(possibly very powerful) to back video out of audio in that way.


The presentation of this project alone is a visual tour de force to say nothing of the technical quality. Beautiful and easily digestible post. As with any interesting, non-toy applied ML problem, the dataset generation is really where the innovation is. It gets a neat little graphic at the end. As far as how the author characterizes the problem, I think the word he's looking for is "semantic segmentation" - he's trying to classify each pixel of the spectrograph as vocal/non-vocal. I'd be curious if he could drop the dataset into pix2pix-style networks and achieve the same results.


Question: has any progress been made in removing reverb?

There are many, many historical recordings (and modern ones made in less-than-ideal circumstances) that suffer badly from reverb. Seems like a valuable use-case that -ought- to be in reach today.


I don't have an answer for this. However, since they do have effective blur reduction/elimination techniques for visual images, I imagine that with enough resources we are not far from reverb/echo reduction in audio.


Blur elimination is usually done with unsharp mask, which works by blurring raster even more and comparing to the original. The output makes the edges more sharp but some information is lost anyway.

Reverb elimination can be done without losses, just with distortions depending on the implementation. To do that, one have to recover cepstral[1] coefficients (with NN) and feed them to spectral filters (no NN needed).

This is feasible, provided somebody prepares a training data set consisting of lots of pairs (sound, same_sound_with_reverb), where sound would be a voice, instrument, applause, etc. and with a different reverb settings. Very likely you'll have to use enormous sample rates, way beyond 44100, because you're supposed to deal with infinitesimal impulse response... Adds up to hardware requirements.

I feel like I've oversimplified something, but it can be done, just lots of fidgeting with all the training sets and a training process itself.

[1] https://en.wikipedia.org/wiki/Cepstrum


Thanks for that, and for the link. I recalled playing long ago with a software that used convolution to let a user choose from different 'reverb spaces' (e.g. Taj Mahal) to put audio into.

It occured to me that in many/most cases the audio from the source arrives first. So -in some cases- multiple models of the 'verb space' could be constructed/refined to allow filtering. Probably much easier for a lone speaker in a small, geometrically simple room -without- a P.A. Maybe not so easy for a speaker using a P.A. in a cathedral. But the Power of Fourier is mighty.


iZotope RX 7, mentioned in a comment above, has something called de-reverb. I didn't try that particular feature, but I did try the music rebalance feature, and it's pretty impressive.


Thanks. I'd call that progress. (Wasn't aware of this 'industry standard audio repair tool'. From WPedia, looks like that just arrived last Sept.)


Just wanted to mention there's some folks doing realtime source separation (not sure exactly how they've implemented it) with a DNN for reduction of background noise in, eg: Skype conversations.

I'm not involved with them in any way, but I've been amazed with its ability to cancel out coffee-shop style noise.

Check out https://krisp.ai/technology/ - Mac/Windows. I wish they had Linux support!

Edit: Appears they don't have Windows support yet.


> Uninterrupted Voice The same krispNet DNN, trained on hundreds of hours of customized data, is able to perform Packet Loss Concealment (predicting lost network packets) for audio and fill out missing voice chunks by eliminating "chopping" in voice calls.

This is both fascinating and horrifying at the same time! I wonder if/when it would be possible to rewrite whole words in real time using a voice that sounds just like you.


Clicked into to the article because I was curious how the training set was created. Using the acapella version is an amazing idea! Wished the article went more in-depth about this section.


Is it possible yet to take a recording of singing and generate a model of the singer for synthesis, like a Vocaloid?


Question: Currently building earphones with great active-noise-cancellation is a secret kept within few companies.

This means they're expensive($300 headphones from Bose etc).

Do neural network make this simpler ?

And do you think they can be applied cheaply enough,say for $99 headphones ?

I assume this will sell really well, and justify creating a dedicated chip, with time.


Doing any kind of neural net anything in realtime is usually not possible due to processing power requirements.


Soon enough there will be an AI filter that will take any old hacky, coughing, wheezing singer running around on stage, singing out of tune - and turn it into virtuoso chops. Maybe even derived from their own voice.

Which will give entirely new meaning to 'lip synching'.


I can’t wait.

Seriously, of all instruments, things like vocals are one of the most heartbreaking to work on and learn. Born male and want to sound like a female singer? You’ll never be able to do that. Same applies for women who wish they had male singing voices. It simply isn’t possible (with the rare exception, I guess).

Or maybe you just don’t like your voices timbre. You can take lessons for years and learn to sing on pitch, and you can alter your vocal tone, but you can’t control every aspect that gives your voice it’s unique sound.

I guess sometimes you just have to be happy with what you have.


A fun thing to do with this would be to slurp the lyrics from one song - the beats from another, some other stream from another and remix the “threads” together into something new.

Basically a giant equalizer that allows you to dim or brighten each channel from multiple sources.


This project I've found to be very useful if you want access to something that like what the article describes. http://isse.sourceforge.net


I'd like to try using this kinda thing to build an automated beat saber map. The ability to orchestrate the beats very specifically would make for excellent mappings.

Alas so many projects, too little time!


Sounds pretty good but exhibits the same artifacts/phasing that I've heard with other source separation. Good for forensics etc but I wouldn't use this for music production


There was a similar demo (I think from Google) here on HN sometime last year that was far more impressive. I can't seem to find it though. Anybody know what it was?


Are there any hearing aid manufacturers taking this approach? Quite incredible.




Registration is open for Startup School 2019. Classes start July 22nd.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: