Shazam for singing (opens in new tab)

Fun post but SoundHound has done this pretty much flawlessly for 10 years. It's also built-in in Google Search and on all Android phones. In that context "Shazam for Singing" is an unfortunate title since it implies nothing like that exist and that Shazam is SOTA, when it's the other way around.

No idea how Shazam can be so far behind and still be synonymous with song identification in everyday speech.

I think the main problem is that you often don't know the lyrics if you're trying to find a song. Recently a friend of mine hummed a song to shazam and it did actually work!

Google search does it already: https://blog.google/products/search/hum-to-search. Worked decently enough the few times I tried it.

Good job! I am still searching for a good solution which can identify songs by singing or humming the melody. There are sites (eg. soundhound) but these are not working all the time.. I think soundhound is working withoud machine learning. With the power of AI today, we could create something which works better. Like Shazam for music.

Why that long intro with that unnecessary narrative?

This could have been summarized in a tweet

"Shazam for singing: Used whisper to transcribe what I sing, then used Google to get the name of the song"

Shazam works with a signal processing technique ("fingerprinting" - https://patents.justia.com/assignee/shazam-entertainment-ltd) that aims to re-recognize the song as originally recorded - so its design goal was "Which song is playing here?" or more precisely "Which song from my database is played here (again)?" rather than "Which is the closest match from my database to this new song rendition/'cover'?".

You can imagine it like a sequence of hash codes or shingles (for modeling gaps/pauses to borrow a term from Web page similarity) for subsequent parts of the songs.

Notably, Shazam does not aim to transcribe the lyrics; so the OP's approach may potentially claim some novelty here. In any case, this experiment shows how great large pre-trained neural language models are for rapid prototyping to put something together quickly - perhaps to test feasibility before attempting to develop something better and more bespoke.

In a related note, Apple may be working on a similar service, at least they filed a related patent application: https://patents.google.com/patent/US6990453B2/en

SoundHound advertises singing as a primary identification vector.

https://www.soundhound.com/soundhound/

Midomi.com is another service tailored to identify your renditions of a song.

What about: “Ken Liiii, a dibba dibba dow doogh”?

This is a novel, (somewhat) hand-rolled approach that I like to see on HN. As others mentioned Google's "Hum to Search" exists, but where's the fun in using Google Wheel™ when I could reinvent it? :)

Your next challenge: make it work with humming. (Hard mode: humming by someone with as terrible pitch as me.)

I suppose when the next GPT has audio input, you could ask it directly to identify songs.

It also makes me wonder how well ChatGPT can directly tell you lyrics verbatim, and how that would be yet another legal issue

I thought this was how Shazam worked as well. I hummed a tune and was super disappointed it didn't identify it. Sounds like a business opportunity to me :)

pedestrian brain: ChatGPT 4, what services can be used to identify a song by singing into a phone's microphone?

gigabrain: What python script can I write that posts an audio file I recorded to an AI I found from a Google Search, but via REST, returns the text, and then posts the texts as lyrics to ChatGPT 4 which can't tell anything after the cut off date

highly regarded