Steve -- You are absolutely right and I appreciate your feedback. Let me know if you have any interest being involved with this project.
In response, I should clarify by pointing out that the company who engaged me in this effort was calling for 80% accuracy. Their target audience does not include minors, so children were not a consideration. They are using SIP lines with great quality not far from decent digital and analog lines. We could probably average better than 90% in their environment. However, I'd rather overachieve than restate those numbers, so they work for me. I have another 1-2 hours of work to get the module close to the 85% range in my environment. I know what I need to do, but may need filter help. I've just been trying to find time. With a good clean signal, I'm sure people can get 90% or better. On this next set of tests, I'd also like to collect more information, including age and zip. Figuring out how to do these things without affecting the tests seems to take more time than the code itself. The numbers: the 65% detection rate I stated earlier is increasing and it probably closer to 70% now. Our actual detection rate is higher, but this gives us a idea of how the module will perform in various settings. There are several negative factors affecting results, including the following: * Pos: I'm running against Asterisk 1.4.9 on a old laptop, a crappy SIP line, and Comcast with a substandard SIP service * Noise: We started testing in the bay area, where it has been raining for a week; that background noise threw off many of the tests * Noise: Many failures came from international callers; noisy lines and background noise (like my wife singing) throw it off * Code: The code isn't done yet and there is room for some improvement; the code needs to be fine tuned for some side cases * Play: Many testers were playing with the system; that includes my wife who claims she can sound like a male (why in the world??) Many of the failures came from noisey lines, playing (like my wife pretending to be a male), and international callers, from what I've observed. We've had the following approximate demographics: male 73%+ (including false female) and domestic 90%+ (without hidden numbers). That fits my estimates of the list demographics combined with the peers and friends I asked to test, so the honor system appears to be working (honesty prevails). If anyone needs the data output by this module (test data for the past few days), I'd be happy to share. It includes the date/time, gender detected, whether that's correct according to the tester, the winning ratio, and the energy levels broken up into 25hz bands. If there is additional information that I should record, which would be benefitial to the community, let me know and I will work on incorporating. Justin Newman (yes, this is the real deal, not spam!) nt_jnewman at yahoo.com ________________________________ From: Steve Underwood <[email protected]> To: Asterisk Users Mailing List - Non-Commercial Discussion <[email protected]> Sent: Friday, February 20, 2009 7:27:12 PM Subject: Re: [asterisk-users] Please help test the gender detection module at 575-613-4392 Hi, Asterisk Asterisk wrote: > We've had a 65% success rate across the board (actually 35% > incorrect). I'm working on bringing that up to 85% or better. Good gender recognisers get >90% success on PSTN lines. In restricted contexts they can get up to 98%. These figures are not entirely honest, as they assume no children call. Recognisers get terrible results for boys who's voices have yet to break. Regards, Steve > > ------------------------------------------------------------------------ > *From:* Ira <[email protected]> > *To:* Asterisk Users Mailing List - Non-Commercial Discussion > <[email protected]> > *Sent:* Thursday, February 19, 2009 5:18:07 PM > *Subject:* Re: [asterisk-users] Please help test the gender detection > module at 575-613-4392 > > At 04:23 PM 2/19/2009, you wrote: > >It got my gender correct the two times I tested, even with the TV > >loud in the background. > > It got me wrong twice, but so do about 30% of the people who call. > > Ira > _______________________________________________ -- Bandwidth and Colocation Provided by http://www.api-digital.com -- asterisk-users mailing list To UNSUBSCRIBE or update options visit: http://lists.digium.com/mailman/listinfo/asterisk-users
_______________________________________________ -- Bandwidth and Colocation Provided by http://www.api-digital.com -- asterisk-users mailing list To UNSUBSCRIBE or update options visit: http://lists.digium.com/mailman/listinfo/asterisk-users
