11 comments

[ 5.1 ms ] story [ 43.5 ms ] thread
Hey all,

I've spent the past few weekends building this:

http://audentifi.com

It's a little app to identify the songs used in YouTube videos because I got fed up of finding awesome new music and then not knowing who the artist was.

The premise:

1. Enter a YouTube URL containing unknown music

2. Click 'Identify songs' to get the artists and songs

3. Profit!

Now unfortunately the current version is pretty primitive and doesn't always get anything right - but you'd be pleasantly surprised the times it gets things bang on! I've certainly added some new albums to my collection since...

Here are some cool videos:

- [Not another GoPro edit](http://audentifi.com/identify?video_url=https%3A%2F%2Fwww.yo...)

- [Fireworks From A Drone](http://audentifi.com/identify?video_url=https%3A%2F%2Fwww.yo...)

- [Freeride 41](http://audentifi.com/identify?video_url=https%3A%2F%2Fwww.yo...)

Intuitively...

- Sports edits usually work really well

- Longer, more complex videos don't work so well (something I'm looking to improve)

Any feedback is appreciated - I'd love to hear any improvements or ideas you guys have. I'm also tweaking the backend to use some machine learning algorithms to determine where the tracks start and end - does anyone have some useful resources on the topic?

Really I'm just testing the waters and seeing if it's worth continuing with the project. I can only afford to keep a couple of backend servers running, so if a lot of people are submitting videos then the waiting time can be quite long...

Thanks for your time guys :)

Sweet.. I'm highlighting this on Product Hunt tmrw.

What's a good email to connect with you?

mike (at) nordkril.se. I've been listening to some of your podcasts - great work!
Just checked it out on a random time lapse video from National Geographic and it was spot on. Good work!
* what is your tech stack? * are you using LSH/Minhash? or something else? * where did you get your music database?
The hashing algorithm is provided by Echonest in their Echoprint service and used to compare against their database (hopefully growing after their Spotify acquisition). However before this can be done, the individual tracks have to picked out - I've not had much luck with Echoprint identifying more than one track at a time.

Python, RQ / Redis for the task queue, RethinkDB for the backend, FFmpeg for transcoding and C for DSP (only in testing).

Just tried a YouTube video of Stevie Ray Vaughan (randomly selected) - got an error message :) Havent checked more but its really useful if it works as you say. I like to find music to jam along with the guitar so this is a tool to identify new stuff. Thanks
Thanks for checking it out. It's very temperamental at the moment - something I'm hoping to improve in the future!
On further inspection it looks like the video was blocked on copyright grounds.