/ Project description
Situation
Many channels, one standard for every conversation.
Telekom answers customer requests by phone, in chat and on smart speakers with digital assistants such as Frag Magenta. Each channel has its own rules. The phone has no screen, chat has no voice, and nobody on a smart speaker wants long menus. New requests keep coming, and each one has to be understood, sound good and feel right.
User research and voice UI design
- Designing and specifying voice dialogues and intent structures
- Expert reviews, functional tests before every release
- Usability tests in several rounds and quantitative UX tests with dozens of participants
Result. A standing research and test function that checks dialogues with people before release and ranks improvements.
New channels
- Requirements and conversation design for chat and smart speakers
- Usability tests, including how users and assistant first get to know each other
Result. Dialogues that fit each channel, instead of a chat dialogue that simply gets read aloud.
Conversation design, NLU and speech output
- Conversational AI design from requirements to implementation
- Training data and intents for language understanding, now also LLM-based use cases
- Tuning speech output so numbers, names and addresses sound natural
Result. A well-practised team that combines design, implementation and quality assurance in live operation.
A custom NLU model
- Evaluated alternatives to the existing classifier
- Trained our own model, built an inference server, load-tested it
Result. YANIC, short for “Yet Another Intent Classifier”, recognises what callers want in the voice channel and runs in production.
What we take away
Good voice assistants are made in operation.
A voice assistant is never finished. Every new request brings new phrasings, new misrecognitions and new spots where the voice stumbles. What works is a steady loop of design, testing with real people and refinement, and a team that knows the system's history. We are happy to tell you in person what that loop looks like and how to set it up in your own organisation.


