Speech Prosody 2006

Dresden, Germany
May 2-5, 2006

Time-domain Noise Subtraction Applied in the Analysis of Lombard Speech

Hansjörg Mixdorff (1), Katja Grauwinkel (1), Martti Vainio (2)

(1) Department of Computer Sciences and Media, Berlin University of Applied Sciences, Germany
(2) Department of Speech Sciences, University of Helsinki, Finland

This paper presents results of the comparison between speech produced in silence and speech in noise, also known as Lombard speech. A temporal filtering algorithm was developed which successfully removes the ambient noise from recordings of Lombard speech by locating and subtracting a recording of the noise performed in the same environment. The filtering algorithm yields overall noise attenuation between 15 and 30 dB without distorting the speech signal like spectral filtering approaches. In the subsequent acoustic analyses we examined the effect of varying levels of noise on vowel formants, glottal spectra and intensity. For most vowels we found significant rises in F1 and F2, but little variation in formant bandwidth. The overall rise in intensity between silent and 80 dB babble noise conditions was found to be of 9 dB. With growing effort higher harmonics are boosted by up to 6 dB whereas the average speech rate only drops by 5%. In Lombard speech the standard deviation of phone intensity is reduced.

Full Paper

Bibliographic reference.  Mixdorff, Hansjörg / Grauwinkel, Katja / Vainio, Martti (2006): "Time-domain noise subtraction applied in the analysis of Lombard speech", In SP-2006, paper 097.