by Michael Womeldorph
Quick CVIS background: I started w/ Bell Labs CVIS in 1989 as a member of the Tier 4 Support team. The support team was arguably the folks closest to the customers and use cases as we supported most trials and live systems. A couple years later, AT&T relocated me to Denver to help bring up the domestic technical assistance center (TAC) then later the international support team (ITAC) for CVIS and Contact Center solutions. I personally enjoyed the international support the most as I got to travel globally on AT&T’s expense and learned a ton about how our solutions were viewed/used in far off places (I ended up in international roles for the next 20 years). So just a couple quick international stories about introducing CVIS into other cultures.
Speech Recognition in the UK: We started introducing CVIS speech rec after our solution in the US was working fairly flawlessly (e.g., recognition at 90%+). The UK seemed like a good market for us, since we had the English version working well. When we first started testing in the lab the accuracy was pretty good (around 85% or so) so we ended up doing a live trial. After starting the live trial, the initial results were quite bad (less than 50% accuracy). It was puzzling as to why the live trial was so poor, but since the system captured and saved the input from the customers, we pulled them down and started listening to the inputs. Funnily enough, upon hearing the inputs, it was clear what the problem was; when prompted to ‘say “one” for X, “two” for Y’ etc., the Brits where swearing at the prompts (e.g., ‘*&%#, “one”). In general, our British friends are much more comfortable swearing in everyday conversations than we are (I learned this first-hand a few years later as I took a 2 year assignment in the UK). At any rate, our genius R&D team came up with the perfect solution; they built speech models for the most commonly used swear words, and changed the script to disregard any of the swear words and to look for the next word. The accuracy of the trial immediately moved up to the +80% range!
Spanish Text-to-Speech and Voice Prompts: When we developed the initial systems into Spanish, someone from the design team and I reached our to our team in Mexico to demonstrate the solutions and see if we could start lining up some trials. We got to the office in Mexico City, set up the demo system, and got some of the key staff in the office together for a demonstration. The reaction was immediately and universally both heartening and disheartening: “That’s really cool! But you can never sell that here”. When we asked the obvious ‘why?’, they explained to us that the demo solution used Castilian Spanish, which would never be acceptable in Latin America. Kind of like us trying to use a NYC or southern accent in the rest of the US (Audrey Audix had a beautiful Ohio accent that is acceptable throughout the US). So our key learning from the trip was that for Latin America, we needed to use voice talent from Colombia, which is acceptable throughout the region, and definitely not Castilian (which would be like trying to use BBC English in Macon Georgia)…
Leave a Reply