NaN in AnalyzerResult

[thomasfillon]

The stress test for analyzers added by 2aa0b7434da9dc06 reveals some nasty NaN.

Nan appears in:

• Onset Detection Function
• IRIT Speech 4Hz Modulation Energy Confidence
• AubioPitch in pitch_confidence (temporary fix by 1a35d497e30043 but @piem may provide an better upstream solution through aubio)

[piem]

On 11/03/2014 14:21, Thomas Fillon wrote:

• AubioPitch in pitch_confidence (temporary fix by 1a35d497e30043 but @piem may provide an better upstream solution through aubio)

good catch. I have a fix for that, it should be included in 0.4.2 (0.4.1 was released yesterday: http://aubio.org/news/20140312-1953_aubio_0.4.1).

cheers, piem

[thomasfillon]

Fixed:

• for IRIT Speech 4Hz Modulation Energy Confidence by 4dd82ae828b1cac5c910ecb10810d58a4f36d132 and
• for Onset Detection Function by 7e096b9e8710d38bc424e02e66ee5375701803c3

I will close this one but the temporary fix for aubio_pitch will need to be remove when Aubio 0.4.2 will be released.

[piem]

On 03/19/2014 11:12 AM, Thomas Fillon wrote:

I will close this one but the temporary fix for aubio_pitch will need to be remove when Aubio 0.4.2 will be released.

ok, but then i'll open a new bug to think about it! :-)

[thomasfillon]

ok so feel free to re-open it ;)

[piem]

np.nan_to_num should be removed once piem/aubio#7 gets fixed.

[thomasfillon]

@piem

Hi ! After a closer look at the YIN algorithm, I wonder if the fix using np.nan_to_num in 1a35d497e30043 is relevant.

How pitch_confidence should be interpreted ? If I understand the YIN article, it seems that harmonic signal give small positive values and non-harmonic signal give high positive values ? Is that correct ? If so, turning NaN to 0 seems a bad idea.

[piem]

After a closer look at the YIN algorithm, I wonder if the fix using |np.nan_to_num| in 1a35d49 https://github.com/yomguy/TimeSide/commit/1a35d497e30043 is relevant.

How |pitch_confidence| should be interpreted ?

close to 1. means high confidence in this pitch candidate. close to 0. means no confidence.

If I understand the YIN article, it seems that harmonic signal give small positive values and non-harmonic signal give high positive values ? Is that correct ?

what positive values?

If so, turning NaN to 0 seems a bad idea.

nans are produced when the input signal is full of zeros. pitch = 0 in that case, but confidence goes wild.

so yes, it makes sense set confidence = 0 on silence before piem/aubio#7 gets fixed. i have a fix for that one, but it needs a bit more testing.

cheers, Paul

[thomasfillon]

OK thanks. So the pitch_confidence is not the Cumulative mean normalized difference function (CMNDF) evaluated by YIN ? @LeCoz : Did you know that ?

[LeCoz]

I suspected that, yes... Le 9 juil. 2014 19:10, "Thomas Fillon" notifications@github.com a écrit :

OK thanks. So the pitch_confidence is not the Cumulative mean normalized difference function (CMNDF) evaluated by YIN ? @LeCoz https://github.com/LeCoz : Did you know that ?

— Reply to this email directly or view it on GitHub https://github.com/yomguy/TimeSide/issues/44#issuecomment-48504254.

[piem]

hi!

i feel like i missed a part of the discussion here.

the confidence is just that, a scalar giving an indication of how much trust should be given to the current pitch evaluation.

if you need something else, please do let me know!

Paul

[thomasfillon]

close by https://github.com/aubio/aubio/commit/e391790ac75d01f4eb89534f17f0a58ba0273b5a