7th International Conference on Spoken Language Processing
September 16-20, 2002
When voice-enabled remote access to information and services is considered, new sources of distortions -due to modern communication networks- should be taken into account. Focusing on mobile communication systems, speech coding distortion and transmission errors severely affect the performance of automatic speech recognition systems.
In this paper, we propose a filtering technique to mitigate the influence of transmission errors on recognition performance. In particular, considering the temporal evolution of each spectral parameter used for recognition as a temporal signal, we postulate that artificial high frequencies are generated in theses signals due to transmission errors. We suggest filtering them out. The proposed technique is assessed for two different parameterisations (type of spectral parameters): MFCC (Mel Frequency Cepstral Coefficients) and LSP parameters (Line Spectral Pairs). The results show that filtering is clearly effective.
Bibliographic reference. Peláez-Moreno, C. / Gallardo-Antolín, A. / Vicente-Peña, J. / Díaz-de-María, F. (2002): "Influence of transmission errors on ASR systems", In ICSLP-2002, 2217-2220.