We used a method to determine the relation between AI/SII and SNR for fluctuating noise in a way quite similar to the way Bill proposed (Smits and Festen,
2013). We found a different relationship between SII and SNR for two-band speech modulated noise compared to steady-state noise (a 2-dB increase in SNR for fluctuating noise gives almost the same increase in SII as a 1-dB increase in SNR for steady-state noise).
This method works well in lab settings but I think the uncertainties are too large to make reliable predictions for real-life situations. The effects of speaking style, speech material, talker etc on percentage correct (the speech recognition function) are
large and the (backward) calculations give large uncertainties in SNR estimates. In situations with fluctuating noise the use of the steady-state noise relationship between SII and SNR could give large errors. With interfering talkers (informational masking)
things get even more complicated.
But, coming back to the question, normal-hearing listeners can follow most conversations in public spaces with some difficulty (often in situations with multitalker
babble which is very similar to stationairy noise). Thus, SNRs around 0-5 dB SNR sounds plausible...
Smits C, Festen JM. The interpretation of speech reception threshold data in normal-hearing and hearing-impaired listeners: II. Fluctuating noise.
J Acoust Soc Am. 2013 May;133(5):3004-15
dr.ir. J.C.M. (Cas) Smits, klinisch fysicus-audioloog |
Universitair Audiologisch Centrum | VU medisch centrum
Deze e-mail is vertrouwelijk en alleen bestemd voor de geadresseerde. Als deze e-mail niet voor u bestemd is, wordt u verzocht dit aan de afzender te melden, het bericht niet verder te
verspreiden en direct te vernietigen. VU medisch centrum staat niet in voor de juiste en volledige overbrenging van de inhoud van deze e-mail.
I agree with Bill except for one (crucial) assumption: in a cocktail-party situation the noise is not stationary (although, as Jont Allen once suggested, adding
talkers to the babble will make it approach stationarity). So, the 0.5 dB SNR is workable in a broad statistical sense and would have to be adjusted almost on a case-by-case basis.
I think Bill's point, which I agree with, is that the SNR is not determined by the noise in the space as much as it is by people
trying to communicate. The SNR that he estimates is not "in the space", but rather "at the ears of the listener" when the talker is trying to communicate to that particular listener, above the noise. Other people trying to communicate to different listeners
make noise for this one. The SNR is therefore roughly constant, somewhere near 0 dB, almost anywhere that's not too quiet.
For me, it's a little higher, after I make people speak up.
On Fri, Jan 24, 2014 at 1:51 PM, Bill Woods <Bill_Woods@xxxxxxxxxxx> wrote:
Since we’re assuming Andy is asking about speech-to-noise ratio when referring to SNR in public spaces, I wondered if the question couldn’t be answered by back-of-the-envelope
calculations using some assumptions and articulation index theory (AI).
It turned out there was quite a long list of assumptions behind the calculation (see below), but the resultant simple calculation yielded a SNR of 0.5 dB, which is quite in line
with the values other posters have indicated may actually be present. Two other facts became apparent as I did this, however, and further motivate my posting.
First, the long list of assumptions makes apparent the long list of influences on the SNR in such a situation. These influences need to be characterized in any attempt to generalize
from SNR measurements in a given scenario, a fact other posters have alluded to.
Second, it is important to note that much work has been done recently to model or empirically characterize these influences on speech intelligibility, and that, while the models
may require more computational power than found on the back of an envelope, they are no problem to execute on current laptop computers. The implication is that it should not be very difficult to determine a reasonably-accurate distribution of SNRs over wide
variation in the assumed listening scenario without any measurements.
The assumptions are:
Talkers have no hearing loss and no cognitive loss.
The talkers are facing each other and speaking in their “mother tongue”.
We know the percent-correct (%C) targeted by the talkers and it is less than 100% (it’s a challenging environment).
We know the nature of the speech in such a conversation, from a low-context vs. high-context perspective.
The talkers are within their critical distance (i.e., ignore reverberation of talkers’ speech).
We know the long-term spectral shape of speech at the eardrums.
We know the long-term spectral shape of noise at the eardrums.
The noise is stationary.
The noise is diffuse.
The diffuse noise in combination with diotic direct-wave target speech generates the equivalent of an “internal” wideband binaural SNR improvement of ~1.0 dB over monaural
Overall level is not too high (i.e., no “roll-over” effect for intelligibility has occurred).
We’re not including lip reading.
These assumptions allow one to, first, determine the AI needed to achieve the assumed target %C given the assumed type of speech, and, second, determine the SNR required with
the assumed spectral shapes to obtain that AI.
For instance, assuming talkers want 95%C with “unfamiliar sentences” then using the polynomial fits from Sherbecoe and Studebaker (JASA 1990) of the ANSI S3.5-1969 transfer functions
between AI and %C, our talkers would need an AI of 0.45. If we assume the noise and speech have the same long-term spectral shape then the SNR can be determined from (SNR+12)/30=0.45 (staying with the 1969 AI method), yielding SNR = 1.5 dB. Subtracting the
binaural SNR improvement yields 0.5 dB.
From: AUDITORY - Research in Auditory Perception [mailto:AUDITORY@xxxxxxxxxxxxxxx]
On Behalf Of Andy Sabin
Sent: Wednesday, January 22, 2014 9:53 AM
Can anyone point me to a reference showing SNRs that are typically observed in public spaces (e.g., restaurants, bars ...etc)? I can find this info for overall SPL, but am having a hard time finding it for SNR.