Need for Audio features probe JSON fields explanation

[Sébastien Vitard]

Hi,

I'm trying to play with the Audio features probe but I don't understand what data are returned in JSON:

"{

""diffSecs"":1.0789999961853027,

""l1Norm"":32.25625,

""l2Norm"":52.28364227939748,

""linfNorm"":22.93468988235943,

""mfccs"":

[

79.26603614546039,

6.74008322111781,

1.3429836887293989,

1.6715114584666757,

3.913417320441085,

3.4400956270685072,

1.6356878915289599,

2.7935779038080404,

2.2162465586243854,

0.12430272756951331,

0.39314584015074067,

3.010446482271461

],

""psdAcrossFrequencyBands"":

[

20950344.040772464,

1336069.3463328325,

695116.1195042991,

194519.98152231556

],

""timestamp"":1369134621.992

}"

Indeed, Javadoc is not helpful since it's just an enumeration of JSON fields and the Built-in probes v3 section doesn't mention this probe...

Can you add a little description of these fields in the Built-in probes v3 section, add a Built-in probes section in v4 documentation, or at least, explain them here?

Thanks in advance.

Sébastien Vitard

[Sébastien Vitard]

Sorry, the link to the Javadoc wasn't on the right page...

[adelich]

Hi Sèbastien

Could you tell me, please, what are duration and period values in your audio features probe scan. I'm asking, because if i scan with duration of 1 second, then i have nothing recorded in DB. But duration=2sec is working.

Adel

вторник, 21 мая 2013 г., 14:27:24 UTC+2 пользователь Sébastien Vitard написал:

[Sébastien Vitard]

I'm using this probe for noise level measurement purpose.

So, I take 30s measurements to get the best estimation without drain the battery too much.

[adelich]

I'm capturing also for noise estimation, but for user's actions recognition purpose.
Doing that i, capture 1 second of Audio Features every 5 seconds.
This means that i set Duration=1s and Period=5s.
The problem is, that if i set duration only 1 sec, then i have not the probe values in DB.
For duration =2 sec it works.

четверг, 23 мая 2013 г., 13:57:36 UTC+2 пользователь Sébastien Vitard написал:

[Sébastien Vitard]

After some search on the Web, I got my answer!

To compute the sound level, we have to apply this formula: 10 * log10( l1Norm^2 ).

Le mardi 21 mai 2013 14:15:20 UTC+2, Sébastien Vitard a écrit :

[Михаил Сысоев]

Hi, could you please provide a link, where you found this formula?

Thank you, Mike

[sebastie...@gmail.com]

Hi Mike,

I do not have access to my work on Funf anymore because I leaved the company for which I had to use it 2 years ago.

But I have found a link which describes how to compute the sound level from a sample : http://digital.ni.com/public.nsf/allkb/FCE0EC0A6B193A028625722E006DE298.

I adapted the formula to the data provided by the probe and it appears that the calculation in the log is quiet similar to what the l1Norm field is (because a tick of the handleAudioStream method happens about every 1s, see class AudioFeaturesProbe). But you can be more precise if you take into account the diffSecs field.

And for your info, the L1 value is the level exceeded for 1% of the time, so it's the most accurate measure (see http://www.epd.gov.hk/epd/noise_education/web/ENG_EPD_HTML/m2/types_3.html).

Best regards.

[Михаил Сысоев]

Hi Sebastien,

Thank you for your help, this is very useful information. Let's me ask you one more question.
So, you actually changed integral to the l1norm? This is not clear for me.

thank you
Best regards, Mike