6 ms·

I've worked on this problem for some time on a personal project, and I'm pretty convinced you can basically solve this problem without deep learning or AI techn

by cnity 5y ago

I've worked on this problem for some time on a personal project, and I'm pretty convinced you can basically solve this problem without deep learning or AI techniques, and instead use non-negative matrix factorization[0] as a bank of note templates (from their spectrograms). I have a fairly well working proof of concept and the approach is supported by the literature. [0]: https://en.wikipedia.org/wiki/Non-negative_matrix_factorization https://en.wikipedia.org/wiki/Non-negative_matrix_factorizat... _edit_ That said, you'd probably need something more hard-core for extraction from an actual track, so you're probably right.

anonymouse008 5y ago
This is the right approach for song recommendation too — try your approach there and see what happens. If you want help on the business side, reach out.
- cnity 5y ago
  That is very kind of you to offer. I don't know if I'd be able to find the time to dedicate to having this function as an actual product though!
  - anonymouse008 5y ago
    You're already leaps and bounds into a good partnership - realistic expectations. Consider it again when you get downtime, no rush over here, more fun.
- rodw 5y ago
  Can you expand on what you mean by "song recommendations" in this context? Do you mean recommendations like "if you like X you might also like Y"? Assuming the answer is "yes", I'm not sure if I follow the leap from "non-negative matrix factorization as a bank of note templates (from their spectrograms)" to "song recommendations". Very loosely speaking my (not wholly uninformed) interpretation of the "note templates" bit is sorta analogous to DFFT analysis with frequency bins centered around the "regular" notes of the chromatic scale - i.e., the cells of the matrix represent signal strength for the frequencies that correspond to conventional western notes (e.g. integer-valued midi note numbers) or (in aggregated form) octave-independent pitch-classes rather than arbitrary frequencies (that might fall between two conventional notes). It's a very useful representation of the component frequencies that appear in conventional music but at the end of the day it's more or less "here are the notes (or pitch classes) active for this given beat". That is, I can imagine the role this information - essentially the musical score itself, or at least the pitch-specific dimension of that score - could play in a song-recommendation engine, but I'm curious how/why that specific spectrogram-template-based representation is significant. Are you suggesting that that specific representation could be applied to song recommendations in a way that similar polyphonic-pitch-per-beat information derived from some other algorithm (FFT analysis for a hypothetical example) could not? Or maybe I've misinterpreted your comment entirely?
  - cnity 5y ago
    More likely, they mean using non-negative matrix factorization but with a bank of feature vectors instead of note templates. NNMF can be used in a wide variety of domains because it essentially encodes the problem of "this thing is a bit like this thing, a bit like this thing, and a bit like this other thing". If instead of numbers representing intensity at different frequencies (as in the spectrograms), the numbers in each vector of the template bank represent other features (such as listener overlap with other artists/songs, or genre representation across multiple continuous "color" axes) then you can recommend music to a listener based on the similarity to songs in their library to those in the template bank.
    - rodw 5y ago
      Ack'd. That makes more sense. I guess I took the comment too literally.
notfed 5y ago
As an amateur musician I see this as a missing holy grail killer app. I'd love to have it just to pull the chords out of some of my own old recordings that I can't figure out how to repeat. If anyone knows of any apps (even prototypes) that can do this, please provide links.
- cnity 5y ago
  I may have to blow the dust off of my prototype.
  - moehm 5y ago
    If you do, make sure to make it a Show HN to let us know!
- cannam 5y ago
  There are various commercial offerings that do this, including Chordify and Capo. Alternatively if you want a flexible free (GPLv2) option with source code, consider the Chordino plugin (http://www.isophonics.net/nnls-chroma http://www.isophonics.net/nnls-chroma, C++ code at https://github.com/c4dm/nnls-chroma https://github.com/c4dm/nnls-chroma) which you can run in Sonic Visualiser (https://www.sonicvisualiser.org/ https://www.sonicvisualiser.org/), or indeed in Audacity. There is a reference describing the Chordino method in the page linked above, but roughly it's not too far from the description of the parent poster - a non-negative least-squares method produces a frame-by-frame semitone-scaled decomposition which is then matched against templates and turned into a chord sequence using a hidden Markov model. Some sort of intelligent smoothing like this is certainly needed, the raw template matching is not especially useful on its own. This type of method is now routinely outperformed by neural networks (see e.g. these MIREX evaluation results from 2018 which compare Chordino with a few other academic methods https://www.music-ir.org/mirex/wiki/2018:Audio_Chord_Estimation_Results https://www.music-ir.org/mirex/wiki/2018:Audio_Chord_Estimat...) but I would suggest that it's still good enough to be useful - and I encourage the parent to continue their work, as it's an interesting area to explore.
  - notfed 5y ago
rodw 5y ago
I've also spent a fair bit of time on this topic and for what it's worth I agree with you. It is a harder problem than the monophonic case (and more sensitive to problems like noise under real-world conditions) but you don't strictly need deep learning or AI techniques to solve it. I mean, computational complexity aside it seems like at least hypothetically you could even just apply basic auto-correlation-style logic to detect the period of the combined wave much like you do in the monophonic case (assuming the chord is sustained for long enough to actually capture that full period, which of course it won't be in the general case). There's nothing magical about a neural-net or other deep-learning-style solution to this problem - at the end of the day that's just an approximation of a formula that could in theory be derived through more direct means anyway. And (as far as I know) there's no reason to believe the polyphonic case is fundamentally resistant to more traditional techniques. And as implied by your comment, the problem is made easier (or at least less resource-intensive) in practice than it is in the abstract: we're mostly interested in audio that's comprised of actual notes from the chromatic scale (rather than a combination of arbitrary frequencies). There's only ~140 or so component frequencies we really need to consider in practice. (Not to mention the semi-predictable repetition/progression patterns you're likely to encounter in most conventional songs. That's inadequate by itself but certainly a good way to error correct, fill in gaps, resolve ambiguous cases, etc.) But that said, it does seem like polyphonic pitch detection is a problem that responds really well to machine-learning techniques. In my experience, even a fairly simplistic ANN (e.g., no hidden layers, ~1k to ~10k weights depending upon how the inputs/outputs are modeled) - when seeded with a little bit of domain-specific knowledge - can very quickly learn to perform reliable polyphonic pitch detection under real-world conditions. To be fair, I haven't quite put my money where my mouth is on this topic (yet): I develop software that includes this sort of functionality and the current production version uses more conventional (or at least direct) analysis rather than so called "deep learning" techniques for polyphonic pitch detection. There are pros and cons to either approach, but I can definitely see why some find the deep learning solution attractive. There's probably some degree of magical thinking involved (i.e., "AI will solve this pattern recognition problem that's too hard for me to work out from first principles"), but it also seems to work really well in this case. For what it's worth I think you've got the right general idea, or at least (based on your brief description) I think I arrived at a solution that's based on some similar concepts and found it fairly effective (beyond the proof-of-concept phase). And as you noted there are related concepts discussed in some of the published academic research. I'd love to hear a little more about your approach if you're willing and able to share any more details. (Noting that at least part of my interest in that topic is selfish, of course.)

by cnity 5y ago

anonymouse008 5y ago

This is the right approach for song recommendation too — try your approach there and see what happens. If you want help on the business side, reach out.

cnity 5y ago
That is very kind of you to offer. I don't know if I'd be able to find the time to dedicate to having this function as an actual product though!
- anonymouse008 5y ago
  You're already leaps and bounds into a good partnership - realistic expectations. Consider it again when you get downtime, no rush over here, more fun.
rodw 5y ago
Can you expand on what you mean by "song recommendations" in this context? Do you mean recommendations like "if you like X you might also like Y"? Assuming the answer is "yes", I'm not sure if I follow the leap from "non-negative matrix factorization as a bank of note templates (from their spectrograms)" to "song recommendations". Very loosely speaking my (not wholly uninformed) interpretation of the "note templates" bit is sorta analogous to DFFT analysis with frequency bins centered around the "regular" notes of the chromatic scale - i.e., the cells of the matrix represent signal strength for the frequencies that correspond to conventional western notes (e.g. integer-valued midi note numbers) or (in aggregated form) octave-independent pitch-classes rather than arbitrary frequencies (that might fall between two conventional notes). It's a very useful representation of the component frequencies that appear in conventional music but at the end of the day it's more or less "here are the notes (or pitch classes) active for this given beat". That is, I can imagine the role this information - essentially the musical score itself, or at least the pitch-specific dimension of that score - could play in a song-recommendation engine, but I'm curious how/why that specific spectrogram-template-based representation is significant. Are you suggesting that that specific representation could be applied to song recommendations in a way that similar polyphonic-pitch-per-beat information derived from some other algorithm (FFT analysis for a hypothetical example) could not? Or maybe I've misinterpreted your comment entirely?
- cnity 5y ago
  More likely, they mean using non-negative matrix factorization but with a bank of feature vectors instead of note templates. NNMF can be used in a wide variety of domains because it essentially encodes the problem of "this thing is a bit like this thing, a bit like this thing, and a bit like this other thing". If instead of numbers representing intensity at different frequencies (as in the spectrograms), the numbers in each vector of the template bank represent other features (such as listener overlap with other artists/songs, or genre representation across multiple continuous "color" axes) then you can recommend music to a listener based on the similarity to songs in their library to those in the template bank.
  - rodw 5y ago
    Ack'd. That makes more sense. I guess I took the comment too literally.

notfed 5y ago

As an amateur musician I see this as a missing holy grail killer app. I'd love to have it just to pull the chords out of some of my own old recordings that I can't figure out how to repeat. If anyone knows of any apps (even prototypes) that can do this, please provide links.

cnity 5y ago
I may have to blow the dust off of my prototype.
- moehm 5y ago
  If you do, make sure to make it a Show HN to let us know!
cannam 5y ago
There are various commercial offerings that do this, including Chordify and Capo. Alternatively if you want a flexible free (GPLv2) option with source code, consider the Chordino plugin (http://www.isophonics.net/nnls-chroma http://www.isophonics.net/nnls-chroma, C++ code at https://github.com/c4dm/nnls-chroma https://github.com/c4dm/nnls-chroma) which you can run in Sonic Visualiser (https://www.sonicvisualiser.org/ https://www.sonicvisualiser.org/), or indeed in Audacity. There is a reference describing the Chordino method in the page linked above, but roughly it's not too far from the description of the parent poster - a non-negative least-squares method produces a frame-by-frame semitone-scaled decomposition which is then matched against templates and turned into a chord sequence using a hidden Markov model. Some sort of intelligent smoothing like this is certainly needed, the raw template matching is not especially useful on its own. This type of method is now routinely outperformed by neural networks (see e.g. these MIREX evaluation results from 2018 which compare Chordino with a few other academic methods https://www.music-ir.org/mirex/wiki/2018:Audio_Chord_Estimation_Results https://www.music-ir.org/mirex/wiki/2018:Audio_Chord_Estimat...) but I would suggest that it's still good enough to be useful - and I encourage the parent to continue their work, as it's an interesting area to explore.
- notfed 5y ago

rodw 5y ago

I've also spent a fair bit of time on this topic and for what it's worth I agree with you. It is a harder problem than the monophonic case (and more sensitive to problems like noise under real-world conditions) but you don't strictly need deep learning or AI techniques to solve it. I mean, computational complexity aside it seems like at least hypothetically you could even just apply basic auto-correlation-style logic to detect the period of the combined wave much like you do in the monophonic case (assuming the chord is sustained for long enough to actually capture that full period, which of course it won't be in the general case). There's nothing magical about a neural-net or other deep-learning-style solution to this problem - at the end of the day that's just an approximation of a formula that could in theory be derived through more direct means anyway. And (as far as I know) there's no reason to believe the polyphonic case is fundamentally resistant to more traditional techniques. And as implied by your comment, the problem is made easier (or at least less resource-intensive) in practice than it is in the abstract: we're mostly interested in audio that's comprised of actual notes from the chromatic scale (rather than a combination of arbitrary frequencies). There's only ~140 or so component frequencies we really need to consider in practice. (Not to mention the semi-predictable repetition/progression patterns you're likely to encounter in most conventional songs. That's inadequate by itself but certainly a good way to error correct, fill in gaps, resolve ambiguous cases, etc.) But that said, it does seem like polyphonic pitch detection is a problem that responds really well to machine-learning techniques. In my experience, even a fairly simplistic ANN (e.g., no hidden layers, ~1k to ~10k weights depending upon how the inputs/outputs are modeled) - when seeded with a little bit of domain-specific knowledge - can very quickly learn to perform reliable polyphonic pitch detection under real-world conditions. To be fair, I haven't quite put my money where my mouth is on this topic (yet): I develop software that includes this sort of functionality and the current production version uses more conventional (or at least direct) analysis rather than so called "deep learning" techniques for polyphonic pitch detection. There are pros and cons to either approach, but I can definitely see why some find the deep learning solution attractive. There's probably some degree of magical thinking involved (i.e., "AI will solve this pattern recognition problem that's too hard for me to work out from first principles"), but it also seems to work really well in this case. For what it's worth I think you've got the right general idea, or at least (based on your brief description) I think I arrived at a solution that's based on some similar concepts and found it fairly effective (beyond the proof-of-concept phase). And as you noted there are related concepts discussed in some of the published academic research. I'd love to hear a little more about your approach if you're willing and able to share any more details. (Noting that at least part of my interest in that topic is selfish, of course.)