In conversation with Nadine and Michel about voicebots, language and speech technology with artificial intelligence in customer service. What is their background and how did they get into speech technology?
Michel Boedeltje is product owner at Telecats. “When I’m not working, I’m cycling. Scheduled for September 2021: Tour for Life, a fundraising trip from Italy to the Netherlands. 1,300 kilometers in 8 days, with ascents totaling 19,000 meters.”
Nadine Glas heads the Speech Lab at Telecats. She loves to dance and hang out with friends. “Haha, I won’t be suffering in the mountains.”
What is your background and how did you get into speech technology?
Nadine: I am originally from a small village in Drenthe (Diever) and started studying in the big city of Groningen when I was 17. I started out studying Communications, but I found that too vague – or not exact enough, if you like. So I applied for a place on an international master’s program, Language & Communication Technologies, part of which was in Nancy, France. On that course, computer science builds a bridge between applied science and linguistics. After my studies, I worked in Paris first at the LIMSI laboratory, and then at a French company in the field of R&D and computational linguistics. Following this, I gained a PhD in human/computer interaction at Télécom ParisTech. After six years in France I returned to the Netherlands. I knew more French companies than Dutch ones, but if you search for jobs online you quickly discover Telecats. That was a pleasant surprise: it fit exactly.
Michel: My background is slightly less international than Nadine’s, but otherwise there are parallels: raised in Groningen and also looking for a technical course. In Enschede, I did technical computer science. It was only further on in my studies that I became fascinated with language and speech technology; I graduated in ‘human media interaction’. There was a lot of focus on the structure and operation of language and speech, which in turn made the computer science a little less technical and mathematical. In 2003 I did an internship with Arjan van Hessen – everyone who wants to do something with language and speech technology sooner or later crosses his path. His lectures made a big impression on me. Through Arjan I ended up at Telecats.
Nadine: Telecats was looking for someone who could further optimize intent recognition. I started out analyzing the algorithms. I introduced new algorithms for specific applications. This involves situations in which not only the subject, but also very specific word combinations are important for properly determining the subject of a question. At the Consumentenbond (a Dutch consumer association), for example, the distinction between “I’m calling about a bill” or “I’m calling about my bill” is essential. If understanding language requires subtle distinctions, you also need subtle algorithms.
What are your roles, and what challenges do you face in their realization?
Nadine: Within the new Speech Lab at Telecats that I am heading up, we are working on models required for products that Michel is the product owner of. Telecats always provided custom solutions for clients. We will convert our models used in this process into products – specific solutions for sectors, for example.
Michel: We are also going to expand the number of languages in which Telecats works, alongside Dutch, Flemish, English and French, with initially German and Spanish and later other languages as well. Telecats always realized a very high quality with its custom solutions. The challenge now is to standardize the applications we have developed in such a way that we can roll them out in different sectors. Our speech routing solution now works for all kinds of customers, with modules in which we realize the customization – for example, integrating with existing applications and training the solution to handle specific jargon of a company. Another challenge is that we want to make the language and speech technology available to end users in the business, in which respect they can do a lot of the configuration themselves. If you can outline a dialog, you can also set it up in the solution – using drag and drop. Simplicity at the front end and all the complexity at the back end, that is one of our challenges now.
Nadine: So that doesn’t mean we’re going to put the same speech recognizer into everything, or start using generic models. We are aiming to make models easier to deploy in new situations or languages, increasing the speed of implementation. My challenge at Speech Lab is training the software.
Where do you want to be in two to three years’ time?
Michel: We are primarily working hard at the moment on developing our portfolio; for example, products aimed at streamlining front office phone traffic and analyzing at the back end. In this context, for example, we are working hard to generate automated summaries of conversations. The next step is also to unleash smart applications on the intermediate stage. Real time agent assist, for instance. In addition, there is a challenge for the coming years to phase out programming of dialogues and allow them to come about automatically based on the input a system receives. An application can then formulate its own next question in order to achieve a goal.
Nadine: We will continue expanding our language coverage. This is obviously mandatory for our internationalization. We have proved that we can do it well for Dutch (including Flemish), French and English; and we need to continue with other languages like Spanish, German, and certainly others. Each day, we continue to expand what we can understand from voice. What I see coming is big progress not only with respect to speech recognition, but also with processing the sound signal itself (the silences, the pitch etc…) to interpret not only what is said, but how it’s said. This will help brands to go even further into understanding their clients and open a lot of new opportunities.