Show HN: Dictate Me – Make Computers Notate What You Sing | Hacker News Reader

Show HN: Dictate Me – Make Computers Notate What You Sing

dictateme.songminseonk.com

32 pointsroaringsheep13 comments

This is a really neat application! Since seeing the fourier image analysis on HN a couple days ago, I've been looking into in-browser DSP.

On my laptop it felt pretty responsive, but I notice that you have a disclaimer not to sing a note faster than 80bpm. Is that a limitation of browser processing power?

Also, I'm curious what you thought of VexFlow? Were there any drawbacks you noted when using it to render notes, and did it play well with angular?

It is because the application will start skipping notes if you start singing faster- I have aggressively filtered entered frequencies which made the dictation speed a bit slower. VexFlow's notation is a little handful to say the least, I had to tweak it a lot to make it work for me.

Cool, great work! Please do post your changes back to the main project, assuming most of them will be generally useful to VexFlow -- 0xfe/Mohit seems quite active and good about responding to pull requests.

> On my laptop it felt pretty responsive, but I notice that you have a disclaimer not to sing a note faster than 80bpm. Is that a limitation of browser processing power?

Might be a limitation of the algorithm used to distinguish individual notes.

It seems like it has promise, and is a neat idea, though the accuracy seems like it needs a lot of work.

I started off by whistling, then humming, then singing (da da da...) the theme from 1812 Overture and it never got close for me. I also sang a scale and it was not close there either.

I will say though: if this worked well, I would use it a lot! A lot of times I have melodies in my head that I find hard to transcribe, so, this would be great!

Forgive my ignorance - is it possible to make this work for chords (ie play a piano and have it transcribe)? Or does that already exist somewhere and I didn't know? Because I would love to have something like that - play piano, and have an application transcribe what you're playing, perhaps automatically detecting key and meter.

What you've described is a Very Hard problem. :) Dozens of researchers around the world are trying to tackle it, with varying results.

Piano is actually one of the easier instruments to automatically transcribe, in part because the notes decay in a predictable way and there's (generally!) no vibrato or changes in tone once a note has been struck. Unlike, say, a clarinet or a guitar, where pitches can be bent after the initial note onset.

Context: I run Soundslice (http://www.soundslice.com/), an interactive sheet-music site, and I've investigated this deeply over the last few years.

If you have a MIDI keyboard, you can plug it into your computer and use a scorewriter app like Finale or Sibelius to do this.

recent discussion of Chordify, youtab et al

https://news.ycombinator.com/item?id=8203196#up_8204455

If you pick thru this conference, there's a huge amount of info on the DSP involved: http://www.ismir.net/proceedings/index.php?&table_name=paper...

It would probably be easier to write a program to transcribe your piano playing from a silent video aimed at the keyboard, than from an audio recording.

Too bad, singing was the only thing that could get past the NSA's automatic transcription service.

pretty cool. there are a handful of similar things that people were working on like this while in school http://music.cs.northwestern.edu/

Very cool!

Reply on news.ycombinator.com