5. Odeon generates an impulse response for each source in the form of “surround-sound” files.

It comes with some predefined speaker maps, but we used a customised 24-speaker map, for which the locations match our ring (described next).

IntroductionIntroduction

Reproducing room acoustics using a computational Reproducing room acoustics using a computational model allied to a loudspeaker ring.model allied to a loudspeaker ring.

*David McShefferty &Michael A. Akeroyd.

MRC Institute of Hearing Research, Glasgow Royal Infirmary, Glasgow.

Frequen

BSA Short-Papers MeetingUniversity of Cambridge

September 14th – 15th 2006

*David@ihr.gla.ac.uk

In acoustically-complex listening situations, hearing-impaired individuals often report particular difficulties, yet these situations are not easily replicated in the laboratory. A synthetic procedure to recreate the acoustics of “real-world” listening spaces within a small laboratory would be highly desirable. Here we report on our initial studies of one such procedure.

Acoustic results: “Full” roomAcoustic results: “Full” room

We used Bruel & Kjaer’s “Odeon” software to recreate the impulse responses of the I.H.R. seminar room.

4. We used frequency shaping to approximate the speakers used for “Dirac” by altering the overall level of each octave bands.

To get the impulse response for each speaker location, we used Bruel & Kjaer’s “Dirac” software. This plays a click (or an MLS sequence, or a sweeping sine wave) through a given loudspeaker, records the response and analyses it. Here are 3 examples:

The computer model : “Odeon”The computer model : “Odeon”

Unfortunately the 24 surround-sound files made by Odeon cannot be played directly through the 24 speakers; there is an automatic gain control and direction is encoded in an “Ambisonics” format.

We have developed an algorithm for undoing these:

For each time point (sample) in the impulse response calculated by

Odeon:• Compute a vector for each of the 24-channels. • Add-up and generate a resultant vector.• Get its amplitude and angle.• Assign a pulse of that amplitude to the corresponding speaker.

The result is 24 separate impulse responses, one for each speaker. These can be convolved with any sound (e.g. an ASL sentence), giving 24 separate sounds, one for each speaker.

If everything works perfectly the result is the percept of the desired sound at the simulated distance in the simulated room.

Auralization through a ring of 24 speakersAuralization through a ring of 24 speakers

Walls(Brick)

Baffle(Foam)

Doors & lectern(wood)

Windows(Double & single

glazing)

Microphone

Sources

Roof (Plaster tiles)

Floor (Thin Carpet on concrete)

We wanted to recreate the acoustics of the new seminar room at the I.H.R. in Nottingham. We had it set up for an experiment on distance: no chairs etc. and a line of speakers facing away from a listener.

The target room : IHR seminar roomThe target room : IHR seminar room

We then tried to recreate these for all 9 locations using a computer model.

4.4 m

6.9 m 3.0 m 3.0 m

door door

door

window window window window

9-m7-m

5-m3-m

1-mListener /Microphone

To

uch

scre

en8-m

6-m4-m

2-m

lectern

divider

6 m

13 m

Viewpoint

of photo

2 m 4 m 8 m

T60 < 100 ms

2.5 m

24 loudspeakers;1-m radius;15º separation.

4 m

Viewpoint of photo

More acoustic results: Simplified roomsMore acoustic results: Simplified rooms

•No frequency shaping: The frequency shape of the sources / speakers were not altered i.e. they were ‘flat’ across each of the frequency bands. This reduces the overall rate of decay.

4-Metre

4-Metre

4-Metre

4-Metre

4-Metre

4-Metre

4-Metre

4-Metre

To get some insight into the effects of Odeon’s many features we gradually simplified the room and various aspects of the modelling procedure to identify the most important steps.

•No scatter: All of the materials used to model the room were given a scatter co-efficient of zero, resulting in specular reflection only.This changes the overall rate of decay.

•No directivity: The sources / speakers were given “omni” directivity patterns, so the energy from them was distributed evenly in all directions. This gives too much reverberant energy.

•Just a basic room: Using the same dimensions and materials but far simplified. This hardly changes the results.

Directivity plot for 500Hz band

These graphs show the “near-instantaneous power” of 4 locations in the real room (Red) and the corresponding Odeon calculations (Black) (each impulse response has been convolved with a 1 ms window, giving a running measurement of power in dB.) The model reproduces many aspects of the impulse responses, such as the echo near t=15ms and the overall decay, but some of the fine detail is missed: e.g. the echoes at 8 ms (bottom row) and 150 & 170 ms (top row).

3. Odeon allows the design of directivity patterns for sources, so we modelled them by entering values at each octave frequency band. The values came from recordings made in the real seminar room with real speakers.

Band (Hz) 63 125 250 500 1000 2000 4000 8000

E.Q Value (dB) -21 -13 -1 4 0 10 10 15

“Zo

om in

”

2-Metre 4-Metre 6-Metre 8-MetreReal roomOdeon

Real roomOdeon

Real roomOdeon

Real roomOdeon

“Zo

om o

ut”

2-Metre 4-Metre 6-Metre 8-MetreReal roomOdeon

Real roomOdeon

Real roomOdeon

Real roomOdeon

t = 15 ms:echo off back wall

2. All of the impulse responses show an echo at about t =15 ms (see panel). This is the echo off the wall behind the microphone.

The equivalent echo in the Odeon models was accurate to within 1 ms in time but (at worst) about 4 dB in near-instantaneous power (see right panel).

This indicates that the Odeon wall is reflecting slightly too much sound, but, curiously, only for the further source distances.

Real roomOdeon modelInverse-square law*

* Theoretical: what would happen if the back wall reflected all the sound.

Level and the perceptual cues to distanceLevel and the perceptual cues to distance

1. The direct-sound level always reduces at 6 dB per doubling in distance. Our initial tests with Odeon showed that it accurately reproduced this effect, (this was used to undo the automatic gain control).

3A … Odeon gives less energy in the reverberant part of the response for the closer distances;

SummarySummary

The primary cue to auditory distance is overall level, and ― thanks to the inverse-square law ― Odeon simulates this generally fine. But a secondary cue to distance is the reverberation relative to the direct sound, which is often experimentally quantified as the direct-to-reverberant ratio. The errors in this measure may therefore be important in some situations.

3B … the direct-to-reverberant ratios mirror this effect;

3. This error, and others in the fine-detail of the response, lead to mismatches in various summary statistics:

3C … and so the overall level of the impulse responses come out too-low in Odeon.

Real roomOdeon modelInverse-square law

Real roomOdeon model

Real roomOdeon model

* Theoretical: what would happen if the reverberation was infinitely weak

A complicating factor ?A complicating factor ?

This figure shows the first few ms of the impulse responses in the real room for all 9 locations (they have been time-aligned/rescaled to the maximum point).

The responses are remarkably similar up-to about t = 0.5 ms; only after that do they differ across location.

The early part of the responses may represent some fundamental limit in our measurement system (probably the loudspeakers?)

This is the corresponding illustration for the Odeon model: the early response is quite different, as it is synthetic and so does not have the real-room limits.

This difference underlies the poor synthesis of the early sound in the closer distances (see “2-m” graph earlier), and will also affect the quantitative measures of accuracy.

Real room, all 9 locations

Odeon model, all 9 locations

Odeon is a very impressive software package and many effects can be replicated. The synthesized impulse responses are similar to the measured responses in the major details.

But the fine details of the synthetic impulses suffer, and there are errors in perceptually-relevant statistics such as the direct-to-reverberant ratio. These errors may limit the experiments that we will be able to do in the synthetic environments.

This work attempted to recreate the acoustic properties of a real room with Bruel and Kjaer’s “Odeon” acoustic-design software.

We did this to study the accuracy of the program before using it to synthesize arbitrary spaces, such as streets, for which we do not have the real acoustics but wish to use as experimental situations.

Band(Hz)63125250500

1000200040008000

EQ value(dB)00000000

1. The room geometry is modelled by listing every point and surface of the room in a text file.

: DMCS 3v06 NOTTSROOM6 : to include recessed window frames, windows in doors,

: radiators etc###Const CeilingLevel 2.4Const RoomWidthWide 6Const RoomWidthNarrow 4.4Const OverallLength 13

:Floor levelPt 101 0 -RoomWidthWide/2 0Pt 102 0 2.33 0Pt 1022 0.26 2.33 0Pt 1023 0.26 RoomWidthWide/2 0

Pt 1024 5.18 RoomWidthWide/2 0Pt 1025 5.18 RoomWidthWide/2-0.43 0Pt 1026 5.55 RoomWidthWide/2-0.43 0Pt 1027 5.55 RoomWidthWide/2-0.2 0Pt 1028 7.23 RoomWidthWide/2-0.2 0Pt 1029 7.23 RoomWidthWide/2 0

Pt 103 OverallLength RoomWidthWide/2 0Pt 104 OverallLength -1.5 0

Pt 1042 OverallLength-2.47 -1.4 0Pt 1043 OverallLength-2.47 -1.5 0

2. Each surface is assigned a realistic ‘material’. This defines its acoustic absorbency. We also apply a scatter co-efficient, (see later).

“Specular” reflection(walls are mirrors)

Scattering