• Partner:

  • DFKI

  • TU Berlin

System Overview

In the following we shortly introduce our system architecture and our graphical user interface. Soon to come: a short video about our user simulation in action.

SpeechEval Architecture

The system’s architecture maintains the modular organization of most SDS, consisting of modules for speech recognition (ASR), natural language understanding (NLU), action planning, answer generation (NLG) and text-to-speech synthesis (TTS). Since our simulation uses speech as interface to the SDS a speech recognizer is the first part of the architecture and takes as input the SDS prompt received from the telephone line. Using speech instead of text or intentions as the interface in the simulation has the advantage of being more realistic and more flexible. We also do not have to worry about introducing simulated ASR errors in our output and our experiments show that the synthesized speech the simulation sends to the SDS is similar in recognition rate to human input.

The following figures give a short overview of the SpeechEval-GUI and the visualization options for the underlying knowledge bases.

Extracted User Templates
Successful Simulation
Scripted Dialog Example