usually very important to distinguish within-speaker variation from between-speaker variation.  In this case, however, between-speaker variation was irrelevant, and the focus was on within-speaker variation only.  The investigation was directed towards answering two questions.  Firstly, are there any differences in the human voice when the body is de

also provided the opportunity to code all the data, and enabled the identification of linguistically equivalent data to be used later in the acoustic analysis.  It also provided the opportunity to hear that the "affected" samples were different from the "normal" samples of speech from each pilot.  There were seven samples of speech (six "normal" samples and one "affected" sample) from the pilot who was known to have become hypoxic.  From the other pilot there were sixteen samples (15 normal and one affected), all recorded on the same day.  In order to check that the "normal" samples really were representative of the pilot's "normal" voice, four samples of speech from that pilot were extracted from ATC recordings made several days prior to the incident.
The acoustic analysis was based on the measurement of three parameters: mean fundamental frequency (F0), formant- or F-pattern, and speech/articulation rate.  F0 was measured using pitch tracking of the entire recording of each communication with ATC.  It was necessary, however, to carefully compare the computer generated pitch tracking with both the waveforms and wide-band spectrograms since the speech analysis equipment (in this case CSL) generated a significant number of extraneous pitch measurements.  For example, clicks as the microphone was turned on or off, and air turbulence from fricatives such as /s/ and /f/ all resulted in F0 measurements which were not related to voicing.  If not removed from the data set, these extraneous meas

(Continued on page 6)

"...it is usually very important to distinguish within-speaker variation from between-speaker variation.  In this case, however, between-speaker variation was irrelevant..."

prived of oxygen, and if so what are the acoustic properties of the voice when subject to hypoxia?  Secondly, it was hoped that the analysis would help determine whether or not the second pilot was in fact hypoxic.
In this article I will outline the procedures used in the investigation.  The results are not reported here, firstly because they are still being finalised, but more importantly, because the findings may become part of a coronial inquiry, and therefore are subject to confidentiality at this stage.  This, of course, is one of the usual frustrations in undertaking forensic speech research involving real-life data!
The first step in the investigation involved an auditory analysis of the data.  From this an accurate orthographic transcription of each of the pilots' communications with ATC were produced.  This