The Intelligibility of Shouted Speech

Notes

A cover sheet on the photocopy reads: "Xeroxed from The Proceedings of the Symposium on the Aeromedical Aspects of Radio Communication and Flight Safety, AGARD/NATO Advisory Report 19, pp. 10-1 to 10-13, December 1969 (London)." Authors printed as Donald J. MacLean and A. Michael Noll, Bell Telephone Laboratories, Incorporated, Murray Hill, New Jersey. Paper number 10; body pages numbered 10-1 through 10-4 with figures and tables following.. Self-archived by the author at http://noll.uscannenberg.org/ScannedPapers/Shouted_Speech_1969.pdf — Noll posts his own papers on his site at noll.uscannenberg.org, which is where this copy comes from. Hosting here is with his agreement, given to Bre Pettis directly on 6 October 2026. Bell Telephone Laboratories was repeatedly asked by Federal agencies to recover the content of radio transmissions from aircraft and spacecraft that had crashed, caught fire or gone out of control. Computer methods already tried on those recordings — analysis and synthesis, time stretching, cepstrum pitch determination, noise stripping — had proved ineffective, and the human ear remained the best analyser. This paper reports controlled experiments to find out why such emergency messages are so hard to decipher. For intelligibility, one talker produced the 300 consonant-vowel-consonant words of a modified Fairbanks rhyme test (six lists of fifty, a closed set of six alternatives per trial) both normally and while shouting, with his own sidetone masked by 110 dB SPL white noise and instructions to shout as if to be heard across a football field. Recording was done in the Murray Hill anechoic chamber with condenser microphones at one metre and one centimetre, band-limited 80 Hz to 10 kHz; 36 dB of peak clipping and various noise levels were applied, and four listeners took the test. At 30 dB signal-to-noise, normal speech scored 98.25 per cent and clipped normal speech 95 per cent, while shouted speech scored 85.2 per cent — much higher than expected. Clipping the shouted speech raised it to 97 per cent; spectrograms of the word “heath” suggest clipping restores the consonant-to-vowel ratio. In a second series five volunteers each spoke and shouted ten words and fourteen colleagues tried to identify them. Correct identification fell from 88.5 per cent for normal voices to 47.2 per cent for shouted ones; only the talker with an unusually high normal pitch stayed recognisable. Pitch roughly tripled under extreme shouting and its range narrowed. The authors conclude that poor signal-to-noise ratio, not nonlinear distortion, is the prime cause, and recommend close-talking microphones and a clipping threshold set to act only on shouted speech. Points of record: Written out of Bell Labs' practical work deciphering crash and emergency radio recordings sent in by Federal agencies; At 30 dB signal-to-noise, intelligibility was 98.25% for normal speech, 95% for clipped normal speech and 85.2% for shouted speech; Peak clipping raised shouted-speech intelligibility to 97%, nearly that of unclipped normal speech — clipping helps rather than harms shouting; Talker identification by familiar colleagues fell from 88.5% (normal) to 47.2% (shouted); Fundamental pitch roughly triples during extreme shouting, and the pitch range across talkers narrows; The authors conclude added noise, not nonlinear distortion, is the major cause of poor emergency intelligibility. Illustrated: Listener response sheet for the rhyme test (Fig. 1); Recording arrangement with condenser microphones at 1 metre and 1 centimetre in the anechoic chamber (Fig. 2); Headphone listening arrangement (Fig. 3); Intelligibility curves for normal speech at 30, 0 and -10 dB signal-to-noise (Fig. 4); Intelligibility curves for shouted speech (Fig. 5); Spectrograms of the word "heath" normal, shouted and clipped-shouted (Fig. 6); Percentage of talkers correctly identified, normal versus shouted (Fig. 7); Pitch frequencies of the five talkers and spectrograms of the word "tan"; Confusion matrices for talker identification, normal and shouted (Tables I and II). People named: Donald J. MacLean (first-named author, Bell Telephone Laboratories), A. Michael Noll (co-author, Bell Telephone Laboratories; also one of the five recorded talkers, initials AMN), J. C. R. Licklider and Irwin Pollack (cited for earlier work on infinite peak clipping and intelligibility), A. S. House, C. E. Williams, M. H. L. Hecker and K. D. Kryter (cited for the closed-response-set articulation test method used). Noted: The supplied citation credits Noll first ("with D. J. MacLean"); the document prints Donald J. MacLean as first author. Figure numbering in the document is inconsistent: both the pitch-frequency chart and the five-talker spectrograms of "tan" are called Fig. 9, and a Fig. 8 is referred to but not described. The microphone type numbers (Brüel & Kjær, read as type 4131 and type 4235) are uncertain in the OCR of this photocopy. The cover sheet gives pp. 10-1 to 10-13; only pages 10-1 to 10-4 carry running text, the rest being figures and tables.