Is this an application of machine learning? Just curious.

[Anand Iyer]

https://mspoweruser.com/google-hum-whistle-or-sing-melody-song/

any idea how it works?   It sounds too good to be true.  Yet to try it though.

[Meenakshi Jayaraman]

Wow.. Yes it works perfectly fine every time.  

On Mon, Oct 19, 2020 at 12:40 PM Anand Iyer <anandd...@gmail.com> wrote:

https://mspoweruser.com/google-hum-whistle-or-sing-melody-song/

any idea how it works?   It sounds too good to be true.  Yet to try it though.

--
You received this message because you are subscribed to the Google Groups "Discussion forum for Computational Thinking" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cs1001-discus...@nptel.iitm.ac.in.
To view this discussion on the web visit https://groups.google.com/a/nptel.iitm.ac.in/d/msgid/cs1001-discuss/0e156285-dc15-4ce8-a627-571d323c4416n%40nptel.iitm.ac.in.

--

Meenakshi Jayaraman

PhD Student,

C/o.Dr.R.Baskar,

Developmental Genetics Lab(314)

Biotechnology Dept.,

IITM

[Himadri Sarkar]

Music search technology has been around for at least a couple of decades, Shazam and SoundHound have monetized it. I found a paper by Shazam engineers describing the algorithm. 

http://www.ee.columbia.edu/~dpwe/papers/Wang03-shazam.pdf

As far as I understood this, they have converted music into a spectrograph and when you sing/hum the computer matches your singing's spectrograph to their database of songs. Essentially they try and narrow down the possibilities. Length of your humming recording and how good of a pitch you can sing should influence the accuracy.

But this is from 15 minutes arm chair research. Please look on the web quite a bit of interesting information in this is available.

 

[Anand Iyer]

Thank you Himadri!

I strongly feel there must've been a lot of changes that came into this field in the last 20 years.  Anyway, will do a bit of looking around...

--

Cheers,

Anand

[Anand Iyer]

https://blog.google/products/search/hum-to-search speaks of machine learning, deep neural networks, a hybrid of spatial partitioning and vector quantization and so on.  

went over my head for the most part, but I noticed a couple of things.  apparently, the fingerprint generation has been simplified over the years.  While there's no mention of machine learning in Shazam's paper, it's mostly only that in this one.  hope to make sense of all this and maybe even be able to contribute to such innovations, at the end of our course!