The standard explanation for the spatial effect of Holophonics is that it is a form of Binaural Recording: two microphones in a real or model head capture the timing, level and filtering cues that the ears use to locate sounds. That explanation leans on the claim that binaural recordings reproduce spatial hearing well. This paper, by Pauli Minnaar, Søren Krarup Olesen, Flemming Christensen and Henrik Møller of the Department of Acoustics at Aalborg University, measures how well. It appeared in the Journal of the Audio Engineering Society, vol. 49, no. 5 (May 2001), pp. 323–336, after presentation at the AES 105th Convention in San Francisco in September 1998. It does not mention Holophonics or Hugo Zuccarelli. Its relevance is that it tests, with controlled data, the kind of recording that the binaural explanation relies on.
What question the study asks
The authors start from earlier Aalborg results, which they summarise. In those studies, listeners who heard recordings made in their own ears located sounds about as well as in real life. Recordings from other people's ears did worse. Recordings from a carefully chosen human head did better than most. And 60% of 30 tested human heads beat even the best artificial head of the time. The new experiment asks whether newer artificial heads have closed that gap.
How the test was run
- Room and sources. A listening room built to IEC 268-13, but without a carpet, held 19 loudspeakers aimed at the listener's head position. Thirteen stood 1 m away in different directions: front, back, left, right and above, with several raised or lowered by 45°. The other six tested distance: straight ahead at 1.7, 2.9 and 5 m, and 45° to the right at 1, 1.7 and 2.9 m.
- Recording heads. Seven artificial heads: the Knowles KEMAR (in its "KEMAR 2" pinna version), the Neumann KU100 (the only one without a torso), the Brüel & Kjær 4100, the Head Acoustics HMS II, the Cortex MK1, Aachen University's ITA and Aalborg's own VALDEMAR, whose pinnae are casts of one human subject. Two human heads, coded AVH and DOL, were picked from the earlier 30-head study. AVH had produced the fewest median-plane errors there, and DOL ranked fourth.
- Playback. Recordings of a 2.2-second female speech sentence and of a 1-second white-noise burst were played back over equalised Sennheiser HE 60 headphones. The equalisation corrected for both the microphone and the headphone.
- Listeners and task. Twenty listeners took part, 10 men and 10 women aged 20 to 30, all with normal hearing and none with earlier experience of these tests. With the loudspeakers in view and their heads held still, they picked which of the 19 loudspeakers the sound seemed to come from. The authors point out that this is a forced choice, not a measure of absolute localisation. Each listener did 20 sessions of 95 stimuli over five days, in an order balanced with a Latin-square design. The whole study gathered 47 500 answers.
The authors sorted mistakes into four types. Out-of-cone errors put the sound on the wrong cone of equal interaural time difference. Within-cone errors confused directions on the same cone. Median-plane errors confused front, above and behind. Distance errors had the right direction but the wrong distance.
What was measured
Speech and noise gave no significant difference in median-plane errors, so the authors pooled them. That gave 3800 responses per condition. Overall error rates were:
| Condition | Median-plane errors | Overall errors |
|---|---|---|
| Real life | 9.7% | 9.1% |
| AVH (human) | 33.0% | 24.9% |
| DOL (human) | 33.2% | 27.8% |
| VALDEMAR | 38.1% | 29.6% |
| ITA | 40.0% | 32.7% |
| KEMAR 2 | 50.5% | 36.2% |
| MK1 | 50.3% | 36.5% |
| HMS II | 53.7% | 38.1% |
| KU100 | 43.0% | 37.8% |
| 4100 | 52.4% | 38.4% |
The paper reports that every recording head did significantly worse than real life, with the directional error categories reaching significance at 0.1% and distance at 5% or better. Left and right were almost never confused. The authors put this down to the difference in arrival time between the two ears, which is unambiguous for sources directly to either side. Most errors came from front/back and elevation confusions, such as a low-front source heard as behind. The KU100 had the most out-of-cone (5.9%) and within-cone (11.7%) errors, which the authors link to its missing shoulders and torso. Distance errors were about the same everywhere, including real life (9.9%). An analysis of variance found no significant difference between heads for distance (p = 0.2124) but a highly significant one for median-plane errors (p < 0.001). It also found large differences between individual listeners.
Listeners improved over the five days. In extra real-life speech sessions, median-plane errors fell from 15.5% on day 1 to 9.6% on day 5, and the curves had not levelled off. The authors give this as the reason a balanced design is needed. Across five related Aalborg experiments, the same head usually scored consistently: only five of 26 within-head comparisons failed to overlap. The authors read this as evidence that localisation with binaural recordings depends mainly on the recording head, provided equalisation is done properly.
What it shows and what it does not
The paper concludes that the artificial heads then available fell short of a well-chosen human head, though some newer heads came close, and that real differences exist between commercial heads. The authors themselves name some limits. A binaural recording fixes the head in the sound field, so the cues a listener gets from moving their head are missing. They also note that the task was a 19-alternative forced choice, which measures identifying a source rather than absolute localisation, and that because the tests took place in a normal listening room rather than an anechoic one, reflections from the room could help listeners localise.
For this dossier, the findings cut both ways:
- Support. Two-channel recordings played over headphones, with nothing added beyond the acoustics of a head and ears, did carry real directional information. Left and right were almost always right, and individual recordings matched real life in the earlier work the paper summarises. This fits the dossier's explanation in Binaural Recording and Head-Related Transfer Function (HRTF) (HRTF).
- Qualification. Recordings made with a head other than the listener's own were clearly worse than real listening, mostly in front/back and elevation. The phrase "comparable to real listening", used in Holophonics Claims vs. Evidence: The Interferometer Theory, the Demonstrations and the Binaural Explanation, holds best for individual recordings. It holds less well for the generic dummy-head recordings that commercial systems use.
Zuccarelli's head-shaped microphone, described in US4680856 — Process for Forming an Acoustic Monitoring Device, was not one of the heads tested. The paper therefore says nothing directly about how well Holophonics recordings localise, and it does not test the Auditory Interferometer Hypothesis.
Reference
Pauli Minnaar and three Aalborg colleagues, "Localization with Binaural Recordings from Artificial and Human Heads", Journal of the Audio Engineering Society vol. 49 no. 5 (May 2001), pages 323–336. Open copy: https://vbn.aau.dk/ws/portalfiles/portal/227876219/2001_Minnaar_et_al_AES_Journal.pdf
Source notes & attribution
- https://vbn.aau.dk/ws/portalfiles/portal/227876219/2001_Minnaar_et_al_AES_Journal.pdf