I recreated Shazam’s algorithm with Go

“An Industrial-Strength Audio Search Algorithm”, the paper where Shazam describes their algorithm (https://www.ee.columbia.edu/~dpwe/papers/Wang03-shazam.pdf) doesn’t have a clear publication date, but https://www.researchgate.net/publication/220723446_An_Indust... indicates it is from 2003.

That patent was filed in the US on 2004-10-21.

IANAL, but to me, that’s a point against that patent in the USA.

I was under the impression that patents aren’t enforceable on open source projects. I assume I am wrong?

I dunno how software patents work but I was under the impression that unless you basically copy paste their code, the courts wouldn't consider it patent infringement as you can't patent the function, but rather the specific thing itself which for software is the exact code itself. But if I'm not understanding something please correct me.

Figure 1 looks interesting since it has both a time and frequency axis, when usually signals have either a time __or__ frequency axis.

Now I'm curious how the Fourier (?) transform of a signal at a __single__ given timepoint is even defined ...

I remember a popular HN post from 10 years ago, that was pulled or the source was pulled because Shazam legally threatened the disclosure of the algorithm. I think it's actually the Google drive file pdf capture from OP's article.

So not enforceable anywhere else than in the U.S.

It's:

(1) deriving a simple fingerprint from the FFT of the audio signal

(2) simple indexing

(3) simple similarity search

You need the signatures of all music on earth for this to work though ;)

Soon!

So you are saying to clone the repo now

Soundhound has always been better than Shazam. It can even pick up people singing and extremely quiet songs

I compared Shazam's, SoundHound's, and BeatFind's recognition library in August 2021. (And tried MusixMatch but it crashed on startup apparently.) Don't think I published it anywhere, these are my raw notes I found among saved chat messages. The format is starting to make more sense now that I'm putting it into a wider window than a chat screen, so I can recommend using a wide browser (94 characters per line should do it). Eyeing the song choices, it looks like I tried to find different genres and artist types (ccMixer/youtube celebrities, to indie, to established) but a larger sample size would obviously have been even better. Still, I hope it's one step up from adding another random opinion!

The conclusion appears to be that BeatFind and Shazam know the most songs, but are also somewhat complementary and all of the services had at least one song they uniquely recognised.

---

    Fun facts:
     * Night Driver (W) said "1 Shazams". I think I was the first person to ever Shazam that. Some of the most obscure things had hundreds, often thousands of shazams!
     * You know where they are taking the hobbits but none of the services do!

    ========

    - ABC = found the song
    - # = number of attempts
    - f = exceptionally fast matching (when it did match, might not be first attempt)
    - ~ = knew one of the songs

    BeatFind:     2B  C     1E    2G 2H  1I  2Jf 2Lf 4M 2Nf 2Of 1Pf 1Rf ~S 2T  1Uf 1Vf 1W 1Xf Y Z
    SoundHound: A 1B        1E 2F 1G 2H  2If     1L         2O  3P  2R     1Tf 2U
    Shazam:       1Bf    1D 1E 1F 1G 1Hf 1I  2J  1L  1M 4N  1O  1Pf 1Rf    1Tf 1Uf 1V  2W 1X  Y Z
    MusixMatch: crashes on startup, presumably it realizes it won't be able to show me ads

    missingno
    Shazam:     A C K Q S
    SoundHound: C D F J K M/N Q S V W X Y Z
    BeatFind:   A D F K Q

    non-universal finds (repeated letter = unique = counts double; slash means same artist so should be counted as one)
    Shazam:     DD F J M/N V X Y Z
    SoundHound: AA F
    BeatFind:   CC J M/N S V X Y Z

    A: Levan Polkka Epic Orchestral Cover version
    B: Pokémon red/blue soundtrack
    C: Mayhem (various songs, it seems either they have all or they have none)
    D: Art Now ft. Snowflake
    E: Hero's Choice
    F: Three Days Grace - Scared
    G: Syrian - Supernova
    H: The Explosion - Here I Am
    I: The Von Bondies - C'mon C'mon
    J: Frank Klepacki - Scouting (C&C TibSun)
    K: Conspiracy - Chaos Theory (demoscene)
    L: Cheshyre - Madness6 (remix) (Newgrounds ID 77998)
    M: Dimrain47 - Twilight Techno
    N: Dimrain47 - Cloud Control
    O: DragonForce - My Spirit Will Go On
    P: Yuki Kajiura - The First Town (SAO)
    Q: THEY'RE TAKING THE HOBBITS TO ISENGARD! THE HOBBITS- THE HOBBITS- TO ISENGARD! TO ISENGARD!
    R: Faithless - Insomnia
    S: Age of Empires 1 soundtrack
    T: Moulin Rouge - El Tango De Roxanne
    U: Van Canto - Master of Puppets
    V: Slack Bird - Jouni
    W: Floppytronic - Night Driver
    X: EgoSalad / Kitboga - Breathe in
    Y: Floppy Drive music: top 4 hits on yt: sweet dreams, imperial march, ghostbusters, beat it. Only ghostbusters was known to any
    Z: Obsidian Shell - Orphanage

Note that what I did not test/review introducing noise (like people talking through it) or filtering (like when you hear the music through a wall)

That's like saying the Hutter prize is useless for anyone who doesn't want highly compressed versions of Wikipedia. The underlying code or algorithm is still interesting to study, use, and remix.

Although, would be curious how good you could get to isolating to a single artist. If you had say one exemplar fingerprint per artist, could an out of dataset fingerprint from their discography cluster to that artist? Obviously not for artists who transitioned musical styles.

Or is the algorithm more feature hash than a clusterable feature vector?

Here we have an open-source algorithm that is useful to anyone with data. It doesn't have to be music

It would actually. But Spotify doesn't allow direct downloads so I had to find the songs on YouTube and download them from there.

You deserve reddit gold for this idea

Why does OP urgently need to change the name?

Notch of Sam?

Not jus' AM?

A. SeekSound vs B. SoundSeek vs C. SoundScout

Which would be a better alternative?

this is a demo of the algorithm, not a full app / hosted service using it with a pre-populated database. The spotify link would be to fingerprint the song and add it to the database

The idea is that you add every Spotify song in the database, and then run your match against them.

Thank you for the time you took to provide such detailed feedback. I really appreciate your honest input. You've raised some valid points that I hadn't really considered.

I agree that the project could definitely use some polishing. I'll prioritize improving the setup instructions and look into adding a file-based DB for flexibility, as well as resolving the npm vulnerabilities. Adding support for directly fingerprinting wav files is a great idea and something I'll prioritize, too.

Regarding the project name, I understand the potential legal implications and will definitely change it. I'd appreciate any suggestions you might have.

I'm excited about the possibility of your contributions. Please, feel free to open a PR whenever you're ready.

Thanks again for your feedback!

You post it on HN to get invaluable comments like yours. Good writeup!

It's to go around the ban of the IP / account by Spotify and to be softer with them, you have to wait between two requests to download songs.

I also use sleep a lot in my code when interfacing with third-party services (multiprocess usually so it's not blocking things, though I'd also totally see myself using a callback pattern or so if the caller can handle those). When it's more than an ad-hoc piece of code, it generally measures how long ago the previous request finished to determine how long to sleep if the next call is made within the cool-off period. If you're not doing that... please don't interface with my server

Does this mean he could accidentally get a $1 million credit card bill from google from someone using his key without his permission? (I don’t know how it works with google.)

Oops... Thank you. I've disabled it.

Adding this to the watch list. Reading this paper was one of the first times I got a 'wow' moment around computing algorthms.

Thanks. I'm glad it inspires you! It'd be awesome to see you take it on. You can clone it and develop it further.

The data comes from a database of fingerprints connected to the server. These fingerprints are created whenever songs are added.

IMHO it’s because audio is „easy“ to manipulate electronically.

You can transform every audio signal into an electronic signal relatively easy - for graphics there is so much more complexity involved just in making them visible.

A speaker that translates electronic signal into sound waves is a super simple contraption at its core.

Edit:/ and audio is striking - it has a profound effect on every human (except deaf of course). If I wanted to demonstrate the power of electronics/computers I would choose audio as well.

I've added that to my to-do list and plan to implement it this weekend.

Thanks! I appreciate the compliment on my Golang; This is actually my first full-fledged project with the language, haha. Feel free to reach out if you have any issues running it.

I think musicbrainz supports this https://musicbrainz.org/doc/AcoustID

Iirc there was some small algorithm, or perhaps even a piece of hardware, that should trigger when music is playing so that the phone isn't active all the time. From there, I guess they could use any old detection algorithm; for me, the magic was in this super-energy-efficient bit of the chain, though I never read up on the details (if they ever provided any)

You mean Suno?