Acoustic-phonetic analysis of normal, loud, and Lombard speech in simulated cockpit conditions

Abstract

It has long been recognized that raising one's voice causes perceptible changes in the speech signal beyond mere increase in overall intensity. This paper reports the results of an extensive analysis of speech produced in a simulated fighter cockpit environment, with helmet and oxygen mask in place. Eight talkers each produced 56 utterances under three conditions: (1) normal, (2) loud (nominally 10 dB above normal), and (3) Lombard (evoked by 90 dB of pink noise played through a headset). A total of 17 671 phonemes were hand marked for analysis using 18 acoustic features (ten frequency bands, spectral COG, low- and high-frequency spectral tilt, F0, F1–3, and duration). For most speakers, both loud and Lombard conditions showed the following shifts in comparison with the normal condition: (1) For vowels and sonorant consonants, lower (0–500 Hz) and higher (5–8 kHz) frequency bands lost energy relative to the mid (1–4 kHz) frequencies; (2) also for vowels and sonorants, F0, F1, and spectral COG all rose; and (3) for fricatives, affricates, and voiceless stops, lower (0–3 kHz) frequencies lost energy relative to higher (4–8 kHz) frequencies. There was much variation in these effects among talkers. [Work supported by Air Force Institute of Technology.]