Harmonic distortion in the TETRA channel and its impact on automatic speech recognition

Since its introduction in the mid 90ies, the Terrestrial Trunked Radio (TETRA) standard for digital radio broadcast has been deployed in most European and Asian government networks and public safety networks. The underlying Adaptive Code-Excitation Linear Prediction scheme employs adaptive code books which are optimized for human speech. However, harmonic distortions of the hardware are amplified as well, which we will show via spectrograms and frequency analysis. While many studies indicate that human perception neglects these distortions up to a Total Harmonic Distortion of about 1%, their impact on Automatic Speech Recognition (ASR) can be dramatic. In this paper, we demonstrate this effect on a strong recognition system for German broadcast news, using a dedicated fire fighter radio transmitter. We will dissect the influence of the hardware and the software components by analysing the word error rate, typical word substitutions and changes in the extracted features. We also suggest countermeasures to obtain a stronger ASR system in this setting.