An Architecture of a Telephone-based System for Speech Data Collection

The work describes the process of building a telephone-based interactive voice response (IVR) system. For the collection of audio data for speech recognition, webbased audio recorders are used so that technical audio equipment does not have to be shipped to the countries with the language in question. However, some parts of the world do not have an Internet infrastructure at all or only a slow one. If these areas provide a telephone infrastructure, an appropriate telephone-based interface is a solution for the collection of audio data. Using existing open source software tools and some classical concepts for implementing conversational systems, we designed and implemented a low cost telephony system for this specific task. Our final system consists of two modules for collecting spontaneous and read speech. Both modules were evaluated with a Wizard of Oz scenario. A second test with the implemented read speech module was performed. As a result from this second evaluation, we can summarize that the test users were able to easily complete their task and are overall satisfied with the system. The system is running and can be accessed via the public switched telephone network on +4972118030681.