Analysing Cockpit and Laboratory Recordings to Determine Fatigue Levels in Pilots’ Voices.

R.Ruiz., LAboratoire de Recherche en Audiovisuel (LA.R.A) , Université de Toulouse, 5 allées Antonio Machado, 31058 Toulouse Cedex 1, France.

C.Legros., Laboratoire d’Acoustique de l’Université de Toulouse-le Mirail (L.A.U.T.M) , Université de Toulouse, 5 allées Antonio Machado, 31058 Toulouse Cedex 1, France.

P.Plantin de Hugues., Bureau Enquêtes et Analyses (B.E.A) pour la sécurité de l’aviation civile, Ministère de l’Ecologie, du Développement et de l’Aménagement Durables, 200 rue de Paris, Zone Sud, Bât 153, Aéroport du Bourget, 93352 Le Bourget Cedex.

Abstract:

Analysis of pilots’ voices was undertaken under controlled laboratory conditions, using aeronautical terminology, standard professional equipment and a standardized recording environment. The aim of the experiment was to determine the way the acoustic characteristics of the voice are modified after a phase of sleep. The results, which are presented in this paper, show very significant variations in the dispersion parameters, like jitters, associated with the fundamental frequency.

In an aircraft cockpit, the electro-acoustic and environmental conditions are not so good. Therefore, it’s essential to improve and adapt voice analysis methods to achieve reliable results. Cockpit Voice Recorder recordings have poor acoustic characteristics due to microphone quality and the signal to noise ratio, in addition to the non-predetermined vocabulary range.

This paper will outline the modifications made to the laboratory methodology and the comparative results obtained, to improve the analysis of CVR recordings.

Comparisons of the test recordings meant that it was feasible to determine, via voice analysis, the state of drowsiness of a pilot and, made possible the study of CVR recordings from accidents using new analytical techniques.

1. Introduction:

Many studies have shown the existence of a link between voice and the emotion of a speaker [1]. The situations which cause emotional disturbance are many and varied such as psychomotor exercises [2], observation of images [3], real or simulated aircraft accidents [4].

In the aeronautical context, voice emotion is searched on Cockpit Voice Recorders (C.V.R) and on recordings made in simulated conditions.

Here the general purpose of the study is to proceed with the same two double approach, but this time for pilot’s fatigue.

First, it aims to ensure that voice changes may occur when a speaker is tired. It is the purpose of the experimentation on sleep inertia described below (paragraph 2). Secondly, voice analysis laboratory methods need to be adapted to Cockpit Voice Recordings (paragraph 3).

Based on this set of results, it becomes possible to study the actual conditions of flight physical fatigue of pilots from the C.V.R and from the analysis of recordings made in the cockpit of the aircraft.

2. Voice and Sleep Inertia: Laboratory Experiment [5] :

Two factors must be taken into account for the study of fatigue due to a workload. The first is the accumulated fatigue during the periods or activity. This will be the focus of a future study. The second is the effects of drowsiness which are presented here.

Pilots who are submitted to daily national rotations can be victims of drowsiness late in the day or even for the early morning flights. The question that arises is whether ability to concentrate and efficiency of their work are always in the best performances.

Medical measures have been done and are completed here by a collection of non-invasive voice data.

2.1 Experimental Conditions:

The experiments were conducted in the service of study on sleep and its disorders of the Créteil Hospital (France) with the assistance of the « Laboratoire d’Anthropologie Appliquée » of the Paris V University.

An airplane pilot is fitted :

-        with electrodes to collect ElectroEncephaloGrams and ElectroCardioGrams and

-        with a nearby microphone mounted on a headband to ensure constant distance between mouth and the microphone. The recordings are made using a DAT recorder.

Three records were held:

Record 1: upon arrival in the experimental room (11h AM)

Record 2: after lunch (14h AM)

Record 3 : The pilot was taken to a room where his sleep is controlled.. After a few minutes of falling asleep he is violently awakened by a very powerful light. It must then perform a number of tasks on a computer in relation to those usually done in flight (15hAM).

The pilot reads the same five sentences in each of the three records.. They belong to aeronautical domain. For example sentence two is : « bravo, victor, charlie montez au niveau deux cinq zéro » that means « bravo, victor, charlie, clim to two five zero level ».

2.2 Voice Analysis:

108 vowels are segmented by only keeping the time period for quasi-stationarity of the signal: 37 for record 1, 30 for record 2 and 41 for record 3.

The analyses are performed using Matlab laboratory programs. The parameters calculated for each vowel are: the mean fundamental frequency <F0> , the associated standard deviation (σ), the Variation Coefficient (V.C), the mean jitter (M.J), the jitter factor (J.F).

  (in %)

  (in Hz)                               (in %)

The most significant results are obtained for the variation coefficient and for the jitter factor as shown in Figures 1 and 2.

Figure 1 : Variation Coefficient of the Fundamental Frequency versus serial number of the vowel.

Figure 2 : Jitter Factor of the Fundamental Frequency versus serial number of the vowel.

Starting from the 67th analysed vowel, ie on awakening of the speaker, the dispersion parameters of the fundamental frequency vary significantly (t-test).

It should be noted that the mean fundamental frequency does not present any visible and significant variation for the Record 3 (Figure 3).

 

figure 3 : Mean Fundamental Frequency versus serial number of the vowel.

Sleep inertia can modify pilots’s voice : an increase of short-term instability of the fundamental frequency is observed.

3. Analysis Methods and Cockpit Voice Recordings [6] :

            3.1 Bandwidth :

The low-frequency limit of the C.V.R bandwidth (150 Hz) prevents detection of the fundamental frequency after a low-pass filtering. Here, the approach is based on the detection of peak amplitudes on the vowel signal. This makes the measurement independant of the low cut-off frequency of the system. The mean fundamental frequency remains almost identical to that of the cepstral analysis.

            3.2 Signal to Noise Ratio :

This parameter is perfectly controlled and high in laboratory conditions. At equivalent speech signal quality, the level of background noise on the CVR tape reduces the signal-to-noise ratio. In most of cases there is no impact on the estimation of parameters related to the fundamental frequency, but spectral estimation is disrupted.

The spectral analysis of the background noise taken from a CVR shows a relatively constant level with frequency between 1000 and 3000 Hz and an important power around 500 Hz (Fig. 4). The spectral slope is about +5dB/octave up to the 2000 Hz octave.

Beyond 500 Hz, the level of background noise is going to strengthen the natural decrease of vowel sound level. The peak at 500 Hz may lead to an overestimation of the sound level around the first formant. Without specific treatment, spectral estimation will be biased.

 

Figure 4 : Background Noise Spectrum (periodogram).

3.3 Modified Spectral Estimation :

The spectral analysis of vowels is performed by an all-pole spectral model obtained from linear predistion (L.PC autocorrelation method). In laboratory conditions pre-emphasis (simple one-zero filter of the form 1-μ.z-1 with μ is near or equal to one) is applied to estimate the transfer function of the vocal without taking into account the effects of the lips radiation and those of the glottal wave. The lack of pre-emphasis leads to an estimate of the spectrum of the vowel.

With the presence of C.V.R background noise , pre-emphasis enhance the spectral effect of noise in for high frequencies. The slope of the frequency response of the vocal tract shows an increase of approximately +11 dB / octave in addition to its normal slope : +6 dB/octave due to pre-emphasis and +5dB/octave due to noise. For the spectrum vowel estimation the increase is about +5 dB / octave leading anyway to excessive high frequency levels (Figure 5).

The recommendation for the study of spectral vowels from CVR is therefore to proceed without pre-emphasis and to perform a first order low-pass filtering before the LPC analysis (Figure 6) .

The analysis conditions are: order 12 for the 8000 Hz sampling frequency, 512 samples in the Hamming window analysis, recovery 50%, pre-emphasis coefficient equal to 0.98 (Figures 5 and 6).

 

Figure 5 : Vowel L.P.C Spectrum (vowel + background noise) without pre-emphasis.

 

 

Figure 6 : Vowel LPC Spectrum (vowel + background noise) without pre-emphasis and with low-pass filtering before sprectrum calculation (Butterworth, cutoff frequency 63 Hz, first order).

 

The background noise is also causing an increase of slightlyless than 20 dB at 500 Hz compared to the frequency band 1300 Hz - 3300 Hz (Figure 4). The low-pass filtering operates a 18 dB decrease of the signal level (vowel + background noise) at 500 Hz from the 63 Hz cutoff  frequency. The low-pass filtering properly reduces the excessive level at 500 Hz. However, it is not an accurate correction because the spectral shape of strengthening around 500 Hz is not taken into account by filtering.

The correction is nevertheless sufficiently efficient to provide a suitable estimation of the vowel spectrum (Figure 6).

4. Conclusion :

            Even if only one pilot voice has been analysed, even if the background noise studied can be modified in others flying situations, the set of results presented in this paper encourage to develop new experiments of voice analysis in real conditions. Sleep inertia experiment indicates that pilot’s voice recordings in the cockpit can provide at least, significant variations of fundamental frequency dispersion parameters. Analysis of C.V.R will also be possible with the spectral corrections advocated.

 

5. References :

            [1] R.Ruiz, C.Legros, A.Guell, “Voice Analysis: Application to the Study of the Influence of a Workload,” J. Acoustique, vol. 3 (2), pp. 153-159 (1990).

            [2] R.Ruiz, E.Absil, B.Harmegnies, C.Legros, D.Poch, “Time- and Spectrum-Related Variabilities in Stressed Speech under Laboratory and Real Conditions,” Speech Com., vol. 20 (1-2), pp. 111-129 (1996).

            [3] R.Ruiz, R.Da Silva Neves, C.Martinot, S.Vautier, “Interactions between Emotional Content of Pictures and Acoustic Features of Speech”, 17th International Congress on Acoustics, Roma, Italy, 6A 14.04 pp 276-277 (2001).

            [4] B.Gramatica, R.Ruiz, C.Legros, “Modifications de la Fréquence Fondamentale de la Voix de Pilotes: Incidents Réels et Simulés,” Journal de Physique IV, colloque C1, vol. 2, pp. 335-338, 2nd Congrès Français d'Acoustique (1992).

            [5] C.Legros, R.Ruiz, “Etude de l’influence d’une période de sommeil sur la voix de pilotes d’avion ”, rapport final de l’étude BEA N°3304/2005, (Mars 2007).

            [6] R.Ruiz, “Prise en compte des spécificités électroacoustiques des communications avion – tour de contrôle pour l’adaptation des mesures de laboratoire sur le signal vocal, phase 2”, rapport final de l’étude BEA N°3841/2007, (Novembre 2007).

Google
  Web auriol.free.fr   


Psychosonique Yogathérapie Psychanalyse & Psychothérapie Dynamique des groupes Eléments Personnels

© Copyright Bernard AURIOL (email : )

November, 01, 2008