Consonant Challenge, Interspeech 2008

Technical details of the collection and postprocessing procedure


Recordings were made in an IAC single-walled acoustically isolated booth at the University of Sheffield. Speech material was collected from a single Bruel & Kjaer (B & K) type 4190 1/2 -in. microphone placed 30 cm in front of the talker. The signal was preamplified by a B & K Nexus model 2690 conditioning amplifier prior to digitisation at 50 kHz by a Tucker-Davis Technologies System 3 RP2.1 processor.

Speakers produced VCVs in isolation by reading out tokens presented on a computer screen, and were given both verbal and written instructions on how to interpret token names, with a particular focus on //, //, /d/, and //. Speakers ran through a practice with the experimenter before speaking alone in the booth. Speakers first produced all VCVs with front stress, followed by end stress. The order of the vowel contexts in which the consonants appeared was: aCa, followed by aCee, aCoo, eeCa, eeCee, eeCoo, ooCa, ooCee, and ooCoo, where a stands for /æ/, ee for /i:/, and oo for /u:/ Collection of speech material was under computer control.

Signals were subsquently post-processed by applying the following steps

  1. High-pass filter to remove frequencies below 50 Hz
  2. Down-sample to 25 kHz (using the MATLAB resample function)
  3. Find endpoints using energy contour
  4. Normalise to have target RMS of 0.05
The resulting files were subsequenly manually analysed (see also the Analysis page). The files containing recoverable endpointing errors were manually endpointed.

File names indicate the gender of the speaker (F or M), the speaker ID (a number between 1 and 24), and the stress pattern: a 1 following the name of the VCV in the file name indicates front stress, a 2 indicates back stress.


Last updated: November 27th, 2007