Re: Objective intelligibility measurements (Christine Rankovic )


Subject: Re: Objective intelligibility measurements
From:    Christine Rankovic  <rankovic@xxxxxxxx>
Date:    Mon, 15 Sep 2008 09:51:44 -0400
List-Archive:<http://lists.mcgill.ca/scripts/wa.exe?LIST=AUDITORY>

Matt: My understanding from your description is that you wish to compare two systems on how they differ with regard to their speech frequency spectra in quiet versus noise, where the speech spectrum differences are caused by a CELP algorithm. I agree with you that the AI will provide useful information for assessing how your systems withstand noise interference, even though, as you already stated, the AIs may be inexact. Christine Rankovic ----- Original Message ----- From: "Matthew Reynolds" <matthew.reynolds@xxxxxxxx> To: <AUDITORY@xxxxxxxx> Sent: Monday, September 15, 2008 5:25 AM Subject: Objective intelligibility measurements Hi all, I am in need of a bit of advice regarding objective measures of intelligibility such as the AI and the STI. I am interested in assessing comparative intelligibility of 2 communication systems in noise with regard to their frequency spectra, however the systems in question use vocoder type codecs to transmit/recreate the speech, introducing non linear distortion that the above mentioned methods do not account for. However, my proposal is this; my signal path in both systems involves CELP algorithms and therefore a direct comparison using AI or STI to assess the intelligibility is not valid, however, it should be possible to gain a score for each system with no additive noise, then repeat the tests with additive noise in the acoustic interface (i.e. in the room with the loudspeaker, so not passing through the electronic signal path, algorithms etc) and see which system 'survives' the noise better with respect to its AI/STI score. This seems to be a valid approach in gaining an idea as to which system will perform better in acoustic noise, although obviously will not produce 'absolute' results for each system. Ideally I'd like to do this with AI as it involves measuring signal to noise ratios in 1/3 octave bands, something which I can readily do, whereas to implement an STI measurement will require purchasing some (probably expensive) software. Thanks in advance for your help Matt Reynolds


This message came from the mail archive
http://www.auditory.org/postings/2008/
maintained by:
DAn Ellis <dpwe@ee.columbia.edu>
Electrical Engineering Dept., Columbia University