Phonetics and Phonology of Speaking Styles: Reduction and Elaboration in Speech Communication Barcelona, Catalonia, Spain September 30 - October 2, 1991

Manual Segmentation and Labelling of Continuous Speech

Knut Kvale (1), Arne Kjell Foldvik (2)

(1) Div. of Electrical Engineering and Computer Science, The Norwegian Institute of Technology, Trondheim, Norway
(2) Dept of Linguistics, University of Trondheim, Dragvoll, Norway

For speech reeearch and development and assessment of speech recognizers and synthesizers segmentation of continuous speech into linguistically defined segments Is desirable. The motivation for studying this problem wae our participation In the European speech research project ESPRIT-SAM In which segmentation and labelling of a large multilingual database Is needed.

Due to the lack of objective criterlons manual segmentation and labelling will exhibit tome Inconsistencies and Is prone to errors. In order to overcome these drawbacks and to make the manual segmentation more consistent and also faster to carry out we developed a set of eegmentatlon rules.

We used the analysis program WAVES on a SUN workstation with the waveform and a broadband spectrogram displayed. Our point of departure was to transcribe what we "heard" and "saw" with phonemic labels and to mark the endpolnt for each phoneme and for pauses.

When comparing segmentation and labelling of the same material performed by the same labeller with a 3 months' Interval the labelling error Is about a half percent and over 96% of the boundaries coincide within ±20 ms. The segmentation and labelling spssd was more than 5 phonemes per minute.

Full Paper

Bibliographic reference.  Kvale, Knut / Foldvik, Arne Kjell (1991): "Manual segmentation and labelling of continuous speech", In PPoSpSt-1991, paper 037.