demo_bpm_extract.py consistently oversamples bpm by +2

[maphouse]

the bpm returned in this function (https://github.com/aubio/aubio/blob/master/python/demos/demo_bpm_extract.py) is consistently +2 bpm above what it should be (I have determined this with hundreds of music files using other bpm detection algorithms and tap notation)

@piem I would appreciate assistance in figuring out what in this function is giving this result and whether it can be fixed!

thanks!

[timvis]

There is also
aubio.tempo().get_bpm()
if I get it right

[maphouse]

There is also aubio.tempo().get_bpm() if I get it right

Thanks for the suggestion! I'm happy to know aubio has a method for bpm, would far prefer to use what you suggest to this other script... Why isn't it anywhere in the docs though? https://aubio.org/manual/latest/py_analysis.html?highlight=tempo#aubio.tempo I can't find get_bpm anywhere...

[timvis]

I found the method by printing dir( aubio.tempo() ).

But it shows only half of the expected BMPs for techno music. Some other music works well. I think it could be a band issue, therefore I would like to filter out unwanted frequency ranges, so I would get only samples for the bass or the high hats. It might not be the right topic here, but maybe someone is reading this... any hints welcome.

[maphouse]

I found the method by printing dir( aubio.tempo() ).

But it shows only half of the expected BMPs for techno music. Some other music works well. I think it could be a band issue, therefore I would like to filter out unwanted frequency ranges, so I would get only samples for the bass or the high hats. It might not be the right topic here, but maybe someone is reading this... any hints welcome.

That approach sounds pretty legit! I'd be curious to know how it improves results... There's def a lot of dance music out there with crazy mids that confuse bpm readers..

With some bpm detection algorithms, you can set upper and lower bounds, which sometimes forces the analyzer to read say 140 instead of 70. This is what happens for me in foobar2000 with https://github.com/stengerh/foo_bpm
I'd prefer to use aubio though (this is the project I'm integrating aubio with btw https://github.com/maphouse/analyze-tag-songs)

I'll try out get_bpm then and follow up here

[maphouse]

I found the method by printing dir( aubio.tempo() ).

But it shows only half of the expected BMPs for techno music. Some other music works well. I think it could be a band issue, therefore I would like to filter out unwanted frequency ranges, so I would get only samples for the bass or the high hats. It might not be the right topic here, but maybe someone is reading this... any hints welcome.

I just got back to this today.

I wasn't able to access the get_bpm method from python. Too difficult! It wasn't obvious to me how to create a "tempo" object (https://aubio.org/manual/latest/py_analysis.html) and access the bpm that way...

It was easier for me to use the shell command "aubio tempo songname", so I integrated it into my script that way. Results for house music are pretty accurate! Much moreso than what I got when using https://github.com/aubio/aubio/blob/master/python/demos/demo_bpm_extract.py

Tracks with BPMs under 100 don't do well though.
Since the script retrieves the median detected bpm, I'm wondering if something like the standard deviation could be used to establish how accurate the detected bpm is...