Investigation of Scaler MIDI and audio detection Part II

Test 2 pad, legato, monotimbral

This test was set up to again see the scaler MIDI performance, but also to check audio detection when the clip contained two identical 1 bar adjacent legato chords. Obviously, the attack part of the envelope would mean that this was different from the same chord sustained over two bars.

The sequence was three triads and one 4 note chord, being C E G/ C E G / C E Bb / C E G A. Naively, these are Cmaj / Cmaj / C7 (no5) / C6. (See ‘fb test mid.jpg’)
[Again the files are at . | BTinternot ; download ‘detection-fb’ ]

Scaler interpreted that as shown in ‘fb scaler.jpg’. Again, technically Bbsus2 (b5)/C and Cmaj add13 makes sense though they might not necessarily make the right base for add a performance on that.

BIAB, in the other hand, detected C / C / C7 / C6 from the MIDI.

Rendering the MIDI as an audio file (‘fb test.wav’) detected 1 bar of Cmaj/G. (‘fb scaler wav.jpg’) - obviouslyu incorrect. This has been noticed before this test, but the behaviour is now confirmed with monotimbral legato pads.

deCoda analysed the audio as at ‘fb decoda.jpg’, being 3 bars of C and 1 of Am. It seems to only analyse sounds as triads, but this needs checking.

As a quick summary (to be added to later)
1 Scaler is very good at analytically analysing MIDI
2 It may be that the derived chords are not necessarily the best guide to melodic phrases to to fit the intended tonality.
3 BIAB is less accurate and normally matches 4 notes when 4 of more are present.
4 However, BIAB scored better in later tests, when chords were mixed with melodic content - see Part III

1 Like

I cannot find it…
Ongoing?

Need to wait till he posts it.

Parts III and IV will follow shortly. They are still concerned with monotimbral test pieces. III will cover a case of chords and melodic content, and Iv will cover more complex melody without overt chords.

My conclusion is that there is no one best solution for detecting scales and chords. However, Scaler is technically very good at detection in situations where it is very good - a bit ‘Zen’, but there are things which trip it up. For instance, it appears to be driven by changes, so it ignores the passage of time for something longer than the declared bar length, which adds to the post detection editing…

To borrow from @jamieh " Patience you must have my young Padawan ” — Yoda
:wink:

Like my wife, who cannot watch a Netflix Series until all seasons have been published, so she can binge watch them all at once :wink:

1 Like