This paper focuses on three Czech speech databases recorded in actual and simulated noisy conditions and explores their suitability for LE analysis and modeling. Parameters of Czech SPEECON, CZKCC car database and newly established Czech Lombard Speech Database (CLSD) are compared. All three databases comprise speech recorded in neutral conditions and speech uttered in noise of the moving car. SNR distribution of the recorded channels, speech fundamental frequency, formant positions and bandwidths, phoneme and word length variations and their overall impact on small vocabulary recognizers performance are analyzed. It is shown that all three databases display speech feature changes across the recording conditions. In SPEECON database these variations do not affect simple recognition task performance much, in CZKCC and CLSD significant recognition degradation has been observed. Due to results of the feature analyses, CZKCC recognition seems to be corrupted rather by background noise than by LE, while in CLSD only LE affects the recognition as the overall SNR is high.
Cite as: Boril, H., Pollák, P. (2005) Comparison of three Czech speech databases from the standpoint of Lombard effect appearance. Proc. Applied Spoken Language Interaction in Distributed Environments (ASIDE 2005), paper 28
@inproceedings{boril05_aside,
author={Hynek Boril and Petr Pollák},
title={{Comparison of three Czech speech databases from the standpoint of Lombard effect appearance}},
year=2005,
booktitle={Proc. Applied Spoken Language Interaction in Distributed Environments (ASIDE 2005)},
pages={paper 28}
}