Mood analysis accuracy

victor's Avatar

victor

01 Mar, 2019 04:19 PM

Are you doing your own mood analysis or do you only rely on last.fm?
I find that the analysis can be rather off. I'm attaching a track (shortened because it was 14M) that is "very depressed, bored", but clearly it's a lively fiddle dance tune. Yes, there are no drums, but that's not the same as "bored".

  1. 1 Posted by victor on 01 Mar, 2019 04:20 PM

    victor's Avatar

    (ok, so the original file went through anyway, it seems.)

  2. Support Staff 2 Posted by hendrik on 02 Mar, 2019 08:38 AM

    hendrik's Avatar

    Hey Victor,

    Are you doing your own mood analysis or do you only rely on last.fm?

    beaTunes 5.x does the following:

    1. The beaTunes database is asked. If a mood is found, it is used and processing stops.
    2. Last.fm is asked. If the track is found, tags are scanned for emotionally significant words and valence/arousal values are inferred. Those are then used and processing is stopped.
    3. As a last resort, AcousticBrainz is asked and its mood values are used.

    The beaTunes database consists of values submitted by users. Those values in turn may have been inferred from Last.fm tags, from AcousticBrainz, or have been manually set. What's currently returned by the beaTunes database server is the median of both the stored valences and arousals, regardless of how they were produced.

    Cheers,

    -hendrik

  3. 3 Posted by victor on 03 Mar, 2019 04:43 PM

    victor's Avatar

    No word on why that one track is mis-characterized?

    Anyway, I was hoping you were doing some fancy DSP’ing that could be tweaked.

    V.

  4. Support Staff 4 Posted by hendrik on 03 Mar, 2019 04:47 PM

    hendrik's Avatar

    No word on why that one track is mis-characterized?

    No, sorry, but it would not change anything.

    Anyway, I was hoping you were doing some fancy DSP’ing that could be tweaked.

    The alternative way to using user content (Last.FM tags) is to do machine learning on user content... This would probably improve results for items that cannot be found on Last.FM.

  5. 5 Posted by victor on 03 Mar, 2019 04:50 PM

    victor's Avatar

    Or plain fancy DSP’ing. You’re already doing some of that to determine the volume. It must be possible to throw more Fourier analysis against it to determine mood and such. ML can play a part in that.

    Victor.

  6. Support Staff 6 Posted by hendrik on 03 Mar, 2019 04:53 PM

    hendrik's Avatar

    The current thinking in the scientific community is supervised learning. As NN input you'd typically use Mel spectrograms. No fancy DSPing necessary. That's the beauty of machine learning.

    I'm totally interested in doing this... but it won't happen for v5. Hopefully v6. We'll see. There is only so much I can do (time is the issue here).

Reply to this discussion

Internal reply

Formatting help / Preview (switch to plain text) No formatting (switch to Markdown)

Attaching KB article:

»

Already uploaded files

  • 01_03_Peter_Man.mp3 13.6 MB
  • 01_03_Peter_short.mov 7.09 MB

Attached Files

You can attach files up to 10MB

If you don't have an account yet, we need to confirm you're human and not a machine trying to post spam.

Keyboard shortcuts

Generic

? Show this help
ESC Blurs the current field

Comment Form

r Focus the comment reply box
^ + ↩ Submit the comment

You can use Command ⌘ instead of Control ^ on Mac