Diagnostic Assessment of Telephone Transmission Impact on ASR Performance and Human-to-Human Speech Quality


Sebastian Möller (Inst. of Communication Acoustics, Ruhr-University Bochum, D-44780 Bochum, Germany)

Ergina Kavallieratou (Wire Communications Lab, University of Patras, GR-26500 Patras, Greece)


SO5: Speech Variabilities & Multilingual ASR


This paper addresses the transmission channel impact on human-to-human speech communication quality as well as on ASR performance. Transmission channels include standard wireline or mobile telephone networks and IP-based networks, which can be operated via different types of user interfaces. In order to gain control over the transmission channel, a simulation model is developed. It implements all types of stationary impairments which can be found in the mentioned networks. Human-to-human speech communication quality in these situations is estimated using a network planning model. Experiments are carried out for assessing ASR performance over the same channel, with three different types of recognizers: two prototypical recognizers used in a telephone-based information server, and a standardized set-up developed under the AURORA framework for distributed ASR. It turns out that some interesting differences exist in behavior between the ASR system performance and speech quality in human-to-human communication. The differences should be taken into account by both developers of ASR systems and transmission network planners. 


Telephone speech, ASR performance, Diagnostic evaluation, Quality prediction, Simulation tools

Full Paper