Raymond Chandler wrote:
What's interesting to me is.... everyone on this thread except you has said that in real-world scenarios, they need the EC for reliability. One of which, does signal processing programming professionally. It seems to me that if you "build a better mouse trap" you must know what's involved in making it work properly. I'm not sure what your background really is, but you'd be hard pressed to match up to Steve's reputation and/or experience.
Public willy-waving is undignified but, in brief, I've built and sold IVRs since 1997, wrote a CAPI-based soft IVR in 1999 (which required software for, inter alia, DTMF detection), developed a software fax modem (V.29, V.27ter, T.30, etc.) which I sold to a CTI card vendor and so on. I've collected some data, of which it is commonly said that the plural of anecdote - which is what we've had so far - is not. The IVR collects a 16 digit DTMF string, terminated by #. TDM->IP conversion was performed by an Asterisk box with an el-cheapo quad E1 card (no EC) for half the calls, and an AS5400 (with EC) for the other half. The proportion of entries missing one or more digit was 3.1% (Asterisk) and 3.3% (AS5400); this is not a statistically significant difference given the sample size. The reason for looking at this criterion is (a) that it's easy to measure, and (b) the most likely way that a DTMF detector will fail in the presence of excess noise, which includes echo, would be to miss a digit. This error rate is the sum of human error + detector error, and I've no measurements to show how this might be split; I would expect it's almost all human. Note that this is a digit error rate of about 1 in 500. This is, of course, only data from one site, but it's a start; it's only by collecting data such as this that one can understand how well one's mouse trap works and whether it needs improvement or not.
That said, it might be a good idea to just agree to disagree as this is starting to sound like the faxing over IP talks I hear a lot. (i.e. "faxing over g.711u with no t.38 works fine for me") Where it might work for some people by some mysterious phenomena, it won't work for the general public. And telling people that they don't need EC, where so many have already said that they obviously do, is just as irresponsible, IMHO, as you claiming Steve was for telling them that they don't need it.
That's a simplification. Simple IVR (record, replay, collect DTMF) probably doesn't need EC; if you're trying to do ASR with barge-in, bridge callers to other callers or operators, etc., then you probably do. I am interested that the recommended solution is 'buy Sangoma' - expensive and proprietary - when Oslec, a FOSS echo cancellers which, by all accounts, works extremely well, is out there and has been for some time. --Dave