Transcript Why do concert halls sound different – and how can we

The Relationship Between Audience
Engagement and Our Ability to Perceive the
Pitch, Timbre, Azimuth and Envelopment of
Multiple Sources
David Griesinger
Consultant
Cambridge MA USA
www.DavidGriesinger.com
Overview
•
This talk consists of an introduction, followed by sections from three fields.
–
–
•
Part one – Physics – describes a physical mechanism by which human
hearing may detect the pitch, timbre, azimuth and distance (near/far) of
several sound sources at the same time, using frequencies the range of
vocal formants (1000Hz to 4000Hz.)
–
–
–
•
The intro states the goal of the talk - to help build better concert halls and opera houses.
To do this we need to understand how acoustics affects the perception of sound.
The acuity of these perceptions is reduced in the presence of reflections and reverberation in
a consistent and predictable fashion.
The consequence of this reduction in acuity is the perception of distance from the source,
and a loss in excitement or engagement with the performance.
A computer model of this mechanism can be used to measure the psychological clarity of a
hall from recordings of live music.
Part two – Psychology – discusses the psychological importance of the
perception of “near” and “far” on the ability of a sound to hold the attention
of the audience.
–
And makes a plea for hall and opera designs that maximize audience engagement
• Part three – Acoustics – looks at the acoustic reasons certain
concert halls are more engaging than others.
– Hall shape does not scale. A shoebox shape that works for a hall with 2000
seats large hall will produce muddy sound over a wide range of seats if it is
scaled to 1000 seats
– Rectangular diffusing elements – coffers and niches – act as frequency
dependent retro-reflectors, and are an essential ingredient in maintaining high
clarity over a large number of seats.
Warning! Radical Concepts Ahead!
•
The critical issue is the amount, the time delay, and the frequency content of
early reflections relative to the direct sound.
•
•
Reflections in the time range of 10 to 100ms reduce clarity, envelopment, and
engagement – whether lateral or not.
–
•
They are typically both early and strong – and interfere with the direct sound.
Side-wall reflections are desirable in the front of a hall, but reduce
engagement in the rear seats.
–
•
and the earliest reflections are the most problematic.
Reflections off the back wall of a stage or shell decrease clarity
–
•
If the direct to reverberant ratio above 700Hz is above a critical threshold, early energy and late
reverberation can enhance the listening experience. But, if not…
They are earlier, and stronger relative to the direct sound in the rear.
Reflections above 700Hz directed into audience sections close to the sound
sources have the effect of reducing the reflected energy in other areas of the
hall – with beneficial results.
–
–
–
These features increase the direct/reverberant ratio in the rear seats
And attenuate the upper frequencies from side-wall reflections in the rear.
Coffers, niches, and/or open ceiling reflectors are invariably present in the best shoebox halls.
A few that work – note the rectangular coffers and niches
Boston Symphony Hall
Vienna Grosse Musikverreinsaal
Amsterdam Concertgebouw
Tanglewood Music Shed
Nice try … – and there are plenty more…
Avery Fisher Hall, New York
Kennedy Center Washington, DC
Alice Tully Hall, New York
Salle Playel, Paris
Introduction
•
This talk is centered on the properties of sound that promote engagement –
the focused attention of a listener.
– Engagement is usually subconscious – and the study of its dependence on
acoustics has been neglected in most acoustic research.
– At some level the phenomenon is well known:
•
•
•
Drama and film directors insist that performance venues be acoustically dry, with excellent speech
clarity and intelligibility.
As do producers and listeners of popular music, and customers of electronically reproduced music of
all genres.
The same acoustic properties that create excitement in a play or film can increase the impact of live
classical music – but many current halls and opera houses are not acoustically engaging in a wide
range of seats.
– Halls with poor engagement decrease audiences for live classical music.
•
Engagement is associated with sonic clarity – but currently there is no
standard method to quantify the acoustic properties that promote it.
– Acoustic measurement s such as “Clarity 80” or C80, were developed to quantify
intelligibility, not engagement.
•
•
Venues often have adequate intelligibility – particularly for music – but poor engagement
Acoustic engineers and architects cannot design better halls and opera
houses without being able to specify and verify the properties they are
looking for.
–
So we desperately need measures for the kind of clarity that leads to engagement.
The story of “near”, “far”, and harmonic coherence
•
The author has been fascinated with engagement for a long time
–
•
This fascination led to a discovery that a major determinant of engagement
was the perception of “near” and “far”,
–
–
•
which humans can determine immediately on hearing a sound, even with only one ear, or
with a single channel of recorded sound.
The perception has vital importance, as it subconsciously determines the amount of attention
we will pay to a sound event.
The importance of this perception, and the speed with which we make it,
argue that determining “near” and “far” is a fundamental property of sound
perception.
–
•
particularly the perception of muddiness in a recording, and the lack of dramatic clarity in a
hall or opera house.
But how do we perceive it, and how can it be measured?
In searching for the answer, the author found that engagement, “near” and
“far”, pitch perception, timbre perception, and direction detection are all
related to the same property of sound:
– the phase coherence of harmonics in the vocal formant range,
~1000Hz to 4000Hz.
Example: The syllables one to ten with four different degrees of phase
coherence. The sound power and spectrum of each group is identical
Near, far, and sound localization
•
The first step to the realization of the fundamental importance of phase
coherence came from author’s listening experience, which suggested that
the perception of “near” and “far” is closely related to the ability to accurately
identify the direction of a sound source.
–
–
–
•
Experiments with several subjects showed that the ability to localize sounds
in a reverberant environment depends on frequencies between 700Hz and
4000Hz,
–
–
•
When individual musicians in a small classical music ensemble sounded engaging and close
to the listener, they could be accurately localized.
And when they sounded distant and non-engaging, they were difficult to localize.
Engagement is mostly sub-conscious and difficult to quantify – but localization experiments
are relatively easy to perform – so I studied localization.
and that poor localization occurs when the sum of early reflections in the time range from
5ms to 100ms from any direction becomes stronger than the direct sound.
The earlier a reflection comes, the larger is its detrimental effect.
With the help of localization data it was possible to construct a measure for
the ability to localize sound in a reverberant environment.
–
The input to the measure is a measured or calculated binaural impulse response at a
particular seat, – ideally with an occupied hall and stage.
Equation for Localizability – 700 to 4000Hz
•
We can use a simple model to derive an equation that gives us a decibel
value for the ease of perceiving the direction of direct sound. The input p(t)
is the sound pressure of the source-side channel of a binaural impulse
response.
–
•
We propose the threshold for localization is 0dB, and clear localization and engagement
occur at a localizability value of +3dB.
Where D is the window width (~ 0.1s), and S is a scale factor:
S is the zero nerve firing line. It is 20dB
below the maximum loudness. POS in the
equation below means ignore the negative
values for the sum of S and the cumulative
log pressure.

 p(t ) dt
S  20  10* log
2
.005
•
Localizability (LOC) in dB =
.005
S  1.5  10* log 
0
p(t ) 2 dt  (1 / D) * 
D .005
0

POS(S  10* log 
.005
p(t ) 2 dt)d
•
The scale factor S and the window width D interact to set the slope of the
threshold as a function of added time delay. The values I have chosen
(100ms and -20dB) fit my personal data. The extra factor of +1.5dB is
added to match my personal thresholds.
•
Further description of this equation is beyond the scope of this talk – but it is
explained on the author’s web-page..
Broadband Speech Data verifies the LOC
equation:
Blue – experimental thresholds for alternating speech with a 1 second reverb time.
Red – the threshold predicted by the localization equation. Black – experimental
thresholds for RT = 2seconds. Cyan – thresholds predicted by the localization
equation.
Measures from live music
•
Binaural impulse responses from occupied halls and stages are very difficult to
obtain!
– But if you can hear something, there must be a way to measure it.
•
Part one of this talk describes a physiologically derived model of human
hearing. The model arose from the search for a measure for “near” and “far”.
– But the model is capable of explaining (and measuring) far more.
– The model provides a means of separating sounds from multiple sources into
independent neural streams,
•
•
And allows independent analysis of each stream for pitch, timbre, and azimuth.
The model may not be neurologically correct in detail
– But it predicts many known properties of human hearing, and shows that all of them
depend on the phase coherence of the incoming sound.
– It provides a method that this phase coherence can be measured from binaural
recordings of live music.
•
This model is the subject of part one of this talk. Parts two and three show why
the model is needed.
Part one – Physics
•
Part one describes a physical mechanism by which human hearing could detect pitch,
timbre, azimuth and distance (near/far) of several sound sources at the same time,
using the phase coherence of harmonics in the range of vocal formants (1000Hz to
4000Hz.)
–
•
Signals from the basilar membrane are analyzed not just for their average amplitude,
but for modulations produced by interference between harmonics.
–
•
–
This this mechanism explains our uncanny abilities to detect the pitch, timbre, azimuth and distance of
several sources at the same time.
And it also predicts the observed decrease in these abilities in the presence of reflections.
The model need not be entirely correct to support the main point of this talk, which is
that the model’s success in predicting what is and is not audible strongly supports the
conclusions of part two and part three.
–
–
–
•
This information derives from the phase relationships between harmonics .
A conceptually simple mechanism is suggested that allows the information from these
modulations to be separated into independent neural streams, one for each sound
source.
–
•
The model is built from functions that are known to be present in human hearing.
The phase coherence of harmonics in the vocal formant range give rise to our abilities to separate sounds
from multiple sources, and independently perceive pitch, timbre, azimuth, and distance for each source.
The acuity of these perceptions is reduced in the presence of reflections and reverberation in a consistent
predictable, and measureable fashion.
In the absence of this acuity sources become psychologically distant and non-engaging.
The model provides a means for measuring the degree of phase coherence – and
thus the engagement – in an individual seat using only live sound as an input.
Perplexing Phenomena
• The frequency selectivity of the basilar membrane is approximately
1/3 octave (~25% or 4 semitones), but musicians routinely hear
pitch differences of a quarter of a semitone (~1.5%).
– Clearly there are additional frequency selective mechanisms in the human ear.
• the fundamentals of musical instruments common in Western music
lie between 60Hz and 800Hz, as do the fundamentals of human
voices.
– But the sensitivity of human hearing is greatest between 500Hz and 4000Hz, as
can be seen from the IEC equal loudness curves.
Blue: 80dB SPL ISO equal
loudness curve.
Red: 60dB equal loudness curve
The peak sensitivity of the ear
lies at about 3kHz.
Why? Is it possible that
important information lies in this
frequency range?
More Perplexing Phenomena
•
Analysis of frequencies above 2kHz would seem to be hindered by the maximum
nerve firing rate of about 1kHz.
–
•
A typical basilar membrane filter above 2kHz has three or more harmonics from each
instrument within its bandwidth
–
•
•
Even in a concert where a string quartet subtends an angle of +-5 degrees or less! (The
ITDs and ILDs are miniscule…)
Why do some concert halls prevent you from hearing several musical lines at once?
–
•
How can we possibly separate them?
How is it possible that in a good hall we can routinely detect the azimuth, pitch, and
timbre of three or more sound sources (musicians) at the same time?
–
•
Why has evolution placed such emphasis on a frequency range that is difficult to analyze
directly?
And what can be done about it?
The hair cells in the basilar membrane respond mainly to negative pressure – they
approximate half-wave rectifiers, which are strongly non-linear devices. How can we
claim to hear distortion at levels below 0.1% ?
Why do many creatures – certainly all mammals – communicate with sounds that
have a defined pitch?
–
Is it possible that pitched sounds have special importance to the separation and analysis of sound?
Answers
•
•
Answers become clear with two basic realizations:
– 1.The phase relationships of harmonics from a complex tone contain
more information about the sound source than the fundamentals.
– 2. And these phase relationships are scrambled by early reflections.
For example: my speaking voice has a fundamental of 125Hz.
– The sound is created by pulses of air when the vocal chords open.
– Which means that exactly once in a fundamental period all the harmonics are in
phase.
•
A typical basilar membrane filter at 2000Hz contains at least 4 of these
harmonics.
– The pressure on the membrane is a maximum when these harmonics are in
phase, and reduces as they drift out of phase.
– The result is a strong amplitude modulation in that band at the fundamental
frequency of the source.
•
When this strong modulation is absent, or noise-like, the sound is perceived
as distant.
Basilar motion at 1600 and 2000Hz
Top trace: A segment of the
motion of the basilar membrane
at 1600Hz when excited by the
word “two”
Bottom trace: The motion of a
2000Hz portion of the membrane
with the same excitation. The
modulation is different because
there are more harmonics in this
band.
In both bands there is strong
amplitude modulation of the
carrier, and the modulation is
largely synchronous.
When we listen to these signals
the fundamental is easily heard
In this example the phases
have been garbled
Nerve firing rates
This picture shows the
amplitude envelope of the
previous picture, plotted in dB.
It represents the rate of nerve
firings from each band. The
rate varies over a sound
pressure range of 20dB.
Nerve cells act like an AM
radio detector, which recovers
the frequency and amplitude of
the modulation, while filtering
away the frequency of the
carrier.
Like the detectors in AM radios, the hair cells (probably) include AGC
(automatic gain control) – with about a 10ms time constant. The response
over short times is linear, but appears logarithmic over longer periods.
AM Radio
• AM radio consists of a carrier at a fixed high frequency that has
been linearly modulated by low frequency signals.
• An AM receiver half-wave rectifies the carrier, and filters out the high
frequency components.
– What remains is the recovered low frequency signals.
– So an AM radio receiver uses a strongly non-linear device to recover a
linear signal.
• But the rectification process can be viewed as a kind of
sampling – it also produces aliases of the modulation.
– In the case of an AM radio the aliases are at very high frequencies, and
can be easily filtered away.
– In the basilar membrane the carrier is close to the frequencies of the
modulation – and the aliases can be problematic.
Amplitude modulation of a noisy carrier
•
The motion of the basilar membrane when excited by phase-coherent
harmonics appears to be an amplitude modulated carrier – but the carrier is
not a fixed frequency, but an artifact of a filter with a finite bandwidth.
– And the frequency of the carrier is within the audio band.
•
Thus, rectification by the hair cells produces aliases that are both broadband and highly audible.
Spectrum of the syllable
“three” from the rectified
and filtered 2000Hz 1/3
octave band (blue) and
the 2500Hz 1/3 octave
band. (red)
Note the fundamental
frequency and its
second harmonic are
the same in both bands.
The garbage is different.
Recovering a linear signal
• To recover a linear signal from these hair cells we need to have to
combine and compare the outputs from many overlapping critical
bands.
– The aliases in each band are different because the carriers have different
frequencies – but the modulations we wish to hear are nearly the same.
• Since for most signals the artifacts are not constant in time – we
must also average the hair-cell firings over a period of time.
– My data suggests an averaging time of 100ms.
• Because the carrier is broad-band, the aliases are also broad-band.
–
The signals are generally narrow band – so broad band signals may be ignored.
• Our hearing mechanism does all of these things.
An amplitude-modulation based basilar
membrane model
A Pitch Detection Model
A neural daisy-chain delays the output of the basilar membrane model by 22us
for each step. Dendrites from summing neurons tap into the line at regular
intervals, with one summing neuron for each fundamental frequency of interest.
Two of these sums are shown – one for a period of 88us, and one for a period of
110us. Each sum produces an independent stream of nerve fluctuations, each
identified by the fundamental pitch of the source.
Pitch acuity – A major triad in two inversions
Solid line - Pitch
detector output for a
major triad – 200Hz,
250Hz, 300Hz
Dotted line – Pitch
detector output for the
same major triad with
the fifth lowered by an
octave: 200Hz, 250Hz
and 150Hz.
Note the high degree of
similarity, the strong
signal at the root
frequency, and the subharmonic at 100Hz
Summary of model
•
•
•
We have used a physiological model of the basilar membrane to convert
sound pressure into demodulated fluctuations in nerve firing rates for a
large number of overlapping (critical) bands.
Our physiological model of the frequency separation mechanism is capable
of analyzing the modulations in each band into perhaps hundreds of
frequency bins.
– Strong, narrow-band signals at particular frequencies are selected for
further processing
The result: we have separated signals from a number of sources into
separate neural streams, each containing the modulations received from
that source.
– These modulations can then be compared across bands to detect
timbre, and IADs and ILDs can be found for each source to determine
azimuth.
Advantages
– The separated streams from each source can be easily analyzed
for timbre, ITD and ILD with known neural circuits.
– The model is conceptually simple – it is built out of a (large) number
of building blocks that are known to exist in human neurology.
• It is easy to see how it could have evolved.
– The circuit is fast. Useful data on timbre, ITD, and ILD is available
within 20ms of the first input.
• As the sound is held pitch and azimuth acuity increases.
– Because the ILD is created by high frequency harmonics, small
differences in azimuth can create large differences in level
• Thus azimuth acuity is high enough to explain our ability to localize musicians.
Speech without reverberation: 1.6kHz-5kHz
Note that the voiced pitches of each syllable are clearly seen. Since the
frequencies are not constant, the peaks are broadened – but the frequency
grid is 0.5%, so you can see that the discrimination is not shabby.
Speech with reverberation: RT=2s, D/R -10dB
The
binaural
audio
sounds
clear and
close.
If we convolve speech with a binaural reverberation of 2 seconds RT, and a
direct/reverberant ratio of -10dB the pitch discrimination is reduced – but still
pretty good!
Speech with reverberation: RT=1s, D/R -10dB
The
binaural
audio
sounds
distant
and
muddy.
When we convolve with a reverberation of 1 seconds RT, and a D/R of -10dB
the brief slides in pitch are no longer audible – although most of the pitches are
still discernable, roughly half the pitch information is lost.
This type of picture could be used as a measure for distance or engagement.
Two violins recorded binaurally, +-15 degrees
azimuth
Left ear - middle phrase
Right ear - middle phrase
Note the huge difference in the ILD of the two violins. Clearly the lower pitched
violin is on the right, the higher on the left. Note also the very clear discrimination
of pitch. The frequency grid is 0.5%
The violins in the left ear – 1s RT D/R -10dB
When we add reverberation typical of a small hall the pitch acuity is reduced
– and the pitches of the lower-pitched violin on the right are nearly gone. But
there is still some discrimination for the higher-pitched violin on the left.
Both violins sound muddy, and the timbre is poor!
Timbre – plotting modulations across critical
bands
• Once sources have been separated by pitch, we can compare the
modulation amplitudes at a particular frequency across each 1/3
octave band, from (perhaps) 500Hz to 5000Hz.
• The result is a map of the timbre of that particular note – that is,
which groups of harmonics or formant bands are most prominent.
• This allows us to distinguish a violin from a viola, or an oboe from a
clarinet.
• I modified my model to select the most prominent frequency in each
10ms time-slice, and map the amplitude in each 1/3 octave band for
that frequency.
• The result is a timbre map as a function of time.
– The mapping works well if there is only one sound source.
Timbre map of the syllables “one two”
All bands show
moderate to high
modulation, and the
differences in the
modulation as a
function of frequency
identify the vowel.
Note the difference
between the “o”
sound and the “u”
sound.
Timbre map of the syllables “one two” with
reverberation 2s RT -10dB D/R
All bands still show
moderate to high
modulation, and the
differences in the
modulation still
identify the vowel.
The difference
between the “o”
sound and the “u”
sound is less clear,
but still
distinguishable.
Timbre map of the syllables “one two” with
reverberation 1s RT -10dB D/R
The clarity of timbre
is nearly gone.
The reverberation
has scrambled
enough bands that it
is becoming difficult
(although still
possible) to
distinguish the
vowels.
A one-second reverberation time creates a greater sense of distance
than a two second reverberation because more of the reflected energy
falls inside the 100ms frequency detection window.
Non-coherent sources
•
So far I have been considering only sources that emit complex tones with a
distinct pitch.
– What about sources that are not coherent, like a modern string section with lots
of vibrato, or pink noise?
•
Nearly any sound source – when band-limited – creates noise-like
modulations in the filtered output.
– Pink noise is no exception. Narrow-band filter it, and the amplitude fluctuates
like crazy.
•
•
•
Sources of this type cannot be separated by frequency into separate
streams – but they can be sharply localized, both by ITD and ILD.
This explains why in a good hall we can easily distinguish the average
azimuth of a string section.
If the strings play without vibrato they are perceived as a single instrument,
with no apparent source width!
Example – Pink noise bursts with identical ILDs
•
I created a signal that consists of a series of pink noise bursts, one of which
is shown below. The noise is sharply high pass filtered at 2kHz.
During the 10ms rise-time
the noise is identical in the
left and right channels.
After 10ms, the noise in the
right channel is delayed by
100us.
The next burst in the series
is identical, but the left and
right channels are swapped.
When you listen to this on
headphones (or speakers)
the sound localizes strongly
left and right.
Azimuth is determined by the ITDs of the modulations – not the onset
Summary of part 1:
•
We have shown that the human ear has evolved to analyze fluctuations or
modulations in the amplitude of the basilar membrane motion at frequencies
above 1000Hz.
– And not necessarily the average amplitude of the motion.
– So long as the phases of the harmonics that create the modulations are not
altered by reflections, the modulations from each source can be separated by
frequency, and separately analyzed for pitch, timbre, azimuth, and distance.
•
The modulations – especially when separated – carry more information
about the sound sources than the fundamental frequencies.
– And allow precise determination of pitch, timbre, and azimuth.
•
•
All of these perceptions depend on the ear’s ability to perceive the direct
sound from the source!!! (And there is currently no standard measure…)
Reflections from any direction – particularly early reflections – scramble
these modulations and create a sense of distance and disengagement.
–
•
But they are only detrimental to music if they are too early, and too strong.
The C language model presented above makes it possible to visualize the
degree to which timbre and pitch can be discerned in a recording of live
music.
– With calibration a single-number measure for engagement should be possible.
Direct sound and Envelopment
•
Recent work by the author in both experiments with several subjects, and in
live lecture demonstrations with loudspeakers, has shown that the sense of
both reverberance and envelopment increases when the direct sound is
separately perceived.
– Where there is no perceivable direct sound the sound can be reverberant, but
comes from the front.
– When the direct sound is above the threshold of localization the reverberation
becomes louder and more spacious.
• Envelopment and reverberance are created by late energy – at least 100ms after the
direct sound.
• When the direct sound is inaudible the brain cannot perceive when a sound has started.
– So effectively the time between the onset of the direct sound and the reverberation is reduced,
and less reverberation is heard.
– In the absence of direct sound syllabic sound sources (speech, woodwinds,
brass, solo instruments of all kinds) are perceived as in front of the listener, even
if reflections come from all around.
• The brain will not allow the perception of a singer (for example) to be perceived as all
around the listener.
• In addition, Barron has shown that reverberation is always stronger in front of a hall
than in the rear – so in most seats sound decays are perceived as frontal.
– But when direct sound is separately perceived, the brain can create two separate
sound streams, one for the direct sound (the foreground) and one for the
reverberation (the background).
• A background sound stream is perceived as both louder and more enveloping than the
reverberation in a single combined sound stream.
Time for Demos
• I will attempt to demonstrate the effects of the direct
sound on the perception of distance, muddiness, and
envelopment.
• The four speakers around the audience will play
reverberation as it might exist in a hall.
• The center channel will produce the direct sound.
• The D/R will vary depending on where you are sitting.
– And I will vary it to give everyone a chance to hear the effects
near the threshold of audibility for the direct sound.
Part 2: Near, Far, and Engagement
•
The apparent closeness of a sound source is a fundamental perception for
all of us.
– We can tell instantly if a person talking is within a few feet of us, or further away
– and this perception has survival value.
– The perception of “Near” depends critically on our ability to perceive the direct
sound – the sound that travels to the listener without reflecting.
– Surprisingly, in a theater or hall it is possible to perceive the performers as both
acoustically close to the listener and enveloped by the hall.
– The best halls (Boston Symphony Hall, Concertgebouw, the front half of the
Musikverrein) provide both, but many, perhaps most, provide only reverberation.
•
Harmonic coherence of speech and music is a principle cue for perceiving
near and far.
– The audio examples in the click box above show the decrease in apparent
distance caused by increasing amounts of harmonic coherence.
– Note that all of the examples have high intelligibility – but their emotional effect is
quite different.
•
The perception of “near” and “far” correlates with ability to localize sound
sources – and the ability to separately hear several musical lines at the
same time.
Amplitude modulation analysis – direct
sound
Spoken syllables “one” to “ten” analyzed by the neural model presented in
Part one. Note the clear detection of the pitch of each vowel
Amplitude modulation analysis “one” to “ten”
with 88ms all-pass reflections
Note the pitch acuity has been reduced, along with the signal to noise ratio.
Amplitude modulation analysis “one” to
“ten” with 133ms reflections
The pitch discrimination with this analysis model (an early one) is
poor with these reflections.
Experiences – Staatsoper Berlin
Barenboim gave Albrecht Krieger
and me 20 minutes to adjust the
LARES system in the Staatsoper.
My initial setting was much to
strong for Barenboim. He
wanted the singers to be
absolutely clear, with the
orchestra rich and full – a
seemingly impossible task.
Adding a filter to reduce the
reverberant level above 500Hz
by 6dB made the sound ideal for
him.
The house continues with this
setting today for every opera.
Ballet uses more of a concert hall
setting – which sounds
amazingly good.
In this example the singers have high
clarity and presence. The orchestra is
rich.
Experiences – Bolshoi – a famously
good hall for opera
The Bolshoi is a large
space with Lots of velvet.
RT is under 1.2 seconds
at 1000Hz, and the
sound is very dry.
Opera here has
enormous dramatic
intensity – the singers
seem to be right in front
of you – even in the back
of the balconies. It is
easy for them to
overpower the orchestra
This mono clip was recorded in the back of the second balcony.
In this clip the orchestra plays the reverberation. The sound
is rich and enveloping
New Bolshoi before modification
The Semperoper was
the primary model
for the design of the
new Bolshoi. As in
Dresden the sound
on the singers is
distant and muddy,
and the orchestra is
too loud.
RT ~1.3 seconds at
1000Hz.
New Bolshoi
Dresden
This theater suffers greatly from having the old Bolshoi next door!
(The theater has received poor reviews from musicains and the press.)
What is it about the
SOUND of this
theater that makes
the singers seem so
far away?
Experiences – Amsterdam Muziektheater
• Peter Lockwood and I spent hours adjusting the
reverberant level using a remote in the hall.
– He taught me to hear the point where the direct sound becomes
no longer perceptible, and the sonic distance dramatically
increases.
– With a 1/2 dB increase in reverberant level, the singer moved
back 3-4 meters.
– In Copenhagen, I once decreased the D/R by one dB while
Michael Schonwandt was conducting a rehearsal. He
immediately waved to me from the pit, and told me to put it back.
• Given a chance to listen A/B, these conductors choose
dramatic intensity over reverberance.
– When they do not have this chance, reverberation is seductive,
and the singers be damned!
Experiences, Copenhagen New
Stage
We were asked to improve
loudness and intelligibility of
the actors in this venue.
64 Genelec 1029s surround
the audience, driven by two
line array microphones, and
the LAREAS early delay
system. A gate was used to
remove reverberation from the
inputs.
5 drama directors listened to a
live performance of Chekhov
with the system on/off every 10
minutes.
The result was unanimous – “it works, we don’t like it.” “The system increases the
distance between the actors and the audience. I would rather the audience did not
hear the words than that this connection is compromised.”
A slide from Asbjørn Krokstad - IoA,NAS Oslo 2008
[With permission]
To succeed:
[in bringing new audience into concert halls…]
ENGAGING
“Interesting” "Nice”
[We need to make the sonic impression of a concert engage the
audience – not just the visual and social perceptions. Especially
since audiences are increasingly accustomed to recordings!]
ENGANGEMENT, not NICE
•
At the IoA conference in Oslo, Asbjørn Krokstad (a musician, conductor,
and Norway’s best-known acoustician) gave a lecture where he insisted that
acousticians needed to provide engagement, not just pleasant music.
– And not just for drama and opera, but for chamber music and symphony too.
– At the end of the lecture he showed a picture of the Teatro Colón in Buenos
Aires, Argentina. “Is this the concert hall of the future” he asked?
• This hall is not a shoebox, but a large semicircular theater with a high ceiling. It ranks
at the top in Beranek’s surveys, and the reverberation time is 1.6 seconds occupied.
• Krokstad may have conducted there.
•
Engagement requires the independent perception of the direct sound
•
We must learn how to provide this essential element in halls.
•
I have been fortunate to hear several of the live broadcasts of the
Metropolitan Opera in a good theater. For example, the performance of
Salome:
– The sound was harsh and dry – radio mikes coupled to directional loudspeakers.
But you could hear every syllable of Mattila’s impeccable German. The
performance was totally gripping!
– This is the dramatic and sonic experience audiences increasingly demand.
Part 3 - Main Points
• The ability to hear the Direct Sound – as measured by LOC or
through binaural recording analysis – is a vital component of the
sound quality in a great hall.
– The ability to separately perceive the direct sound when the D/R is less
than 0dB requires time. When the d/r ratio is low there must be
sufficient time between the arrival of the direct sound and the build-up of
the reverberation if engagement is to be perceived.
• Hall shape does not scale
– Our ability to perceive the direct sound – and thus localization,
engagement, and envelopment - depends on the direct to reverberant
ratio (d/r), and on the rate that reverberation builds up with time.
– Both the direct to reverberant ratio (d/r) and the rate of build-up change
as the hall size scales – but human hearing (and the properties of
music) do not change.
– A hall shape that provides good localization in a high percentage of
2000 seats may produce a much lower percentage of great seats if it is
scaled to 1000 seats.
– And a miniscule number of great seats if it is scaled to 500 seats.
Frequency-dependent diffusing elements are
necessary, and they do not scale.
•
The audibility of direct sound, and thus the perceptions of both localization
and engagement, is frequency dependent. Frequencies above 700Hz are
particularly important.
– Frequency dependent diffusing elements can cause the D/R to vary with
frequency in ways that improve direct sound audibility.
– The best halls (Boston, Amsterdam, Vienna) all have ceiling and side wall
elements with box shape and a depth of ~0.4m.
• These elements tend to send frequencies above 700Hz back toward the orchestra and
the floor, where they are absorbed. (The absorption only occurs in occupied halls – so
the effect will not show up in unoccupied measurements!)
• The result is a lower early and late reverberant level above 700Hz in the rear of the hall.
• This increases the D/R for the rear seats, and improves engagement.
– The LOC equation is sensitive to all reflections in a 100ms window, which will include many
second-order reflections, especially in small halls.
•
Replacing these elements with smooth curves or with smaller size features
does not achieve the same result.
– Some evidence of this effect can be seen in RT and IACC80 measurements
when the hall and stage are occupied.
•
Measurements in Boston Symphony Hall (BSH) above 1000Hz show a clear
double slope that is not visible at 500Hz.
– The hall has high engagement in at least 70% of the seats.
Boston Symphony Hall, occupied hall and stage, stage
middle to front of first balcony, 1000Hz
Note the clear double-slope decay, with the first 12dB decaying at RT = 1s
The direct sound is clearly dominant at this frequency in this seat. The
sound is very good – Leo Beranek’s favorite seat!
Boston Symphony Hall, occupied, stage to front
of balcony, 250Hz
But at 250Hz, there is no evidence of the direct sound. It has
been swamped by reverberation.
We need better measures
• Current acoustic measures ignore both the D/R and the time gap
between the direct (the first wavefront) and the reverberation.
– RT, C80, and EDT all ignore the strength of the direct sound and the
effects of musical style on the audibility of the D/R.
– IACC comes close, but measures only lateral reflections.
• LOC and my hearing model attempt to supply a simple measure for
a basic human perception which depends on direct sound.
– Impulse response measurements under occupied conditions are
notoriously difficult to obtain.
– But this problem can be circumvented through binaural recordings of
live sound.
• The hearing model presented in part one promises to supply
measures that use binaural recordings of actual performances as
inputs.
– And the ability to listen to these recordings to test the validity of these
measures against the true experience.
Why do large halls sound different?
• In Boston Symphony Hall (BSH), and the Amsterdam
Concertgebouw (CG) the reverberation decay is nearly
identical, but the halls sound different.
– The difference can be explained using the same model that was
used to develop LOC.
– Lacking good data with an occupied hall and stage I used a
binaural image-source model with HRTFs measured from my
own eardrums.
Reverberation build-up and decay – from models
Amsterdam
LOC =
+6dB
Boston
LOC =
4.2dB
The seat position in the model has been chosen so that the D/R is -10dB for a continuous note.
The upward dashed curve shows the exponential rise of reverberant energy from a source with
rapid onset and at least 200ms length. The reverberation for the dotted line is exponentially
decaying noise with no time gap. The solid line shows the image-source build up and decay
from a short note of 100ms duration. Note the actual D/R for the short note is only about -6dB.
The initial time gap is less in Boston than Amsterdam, but after about 50ms the curves are
nearly identical. (Without the direct sound they sound identical.) Both halls show a high value
of LOC, but the value in Amsterdam is significantly higher – and the sound is clearer.
Comparisons of C80, C50, IACC80, and LOC
•
Conventional measures for the models of Amsterdam Concertgebouw and Boston
Symphony Hall give the following results:
•
Amsterdam: C80 = .43dB, C50 = -2.8dB, IACC80 = .38, LOC = +6dB
•
BSH:
•
Half-Size BSH: C80 = 3.7, C50 = 1.7, IACC80 = .15, LOC = 0.5dB
•
Only the IACC80 shows that Amsterdam might have more direct sound than Boston.
The standard Clarity measures predict the opposite – and predict that the small hall
would have high clarity, and it does not.
•
But IACC80 is sensitive only to lateral reflections. Strong reflections from the front,
overhead, or rear do not affect IACC.
•
An IACC of 0.22 would usually be considered too low for good sound. In spite of this
BSH has both clarity and good localization in this seat.
C80 = .65dB, C50 = -2.1dB, IACC80 = .22, LOC = +4.2dB
Smaller halls
• What if we build a hall with the shape of BSH, but half
the size?
– The new hall will hold about 600 seats.
– The RT will be half, or about 1 second.
– We would expect the average D/R to be the same. Is it? How
does the new hall sound?
– If the client specifies a 1.7s RT will this make the new hall better,
or worse?
Half-Size Boston
The gap between the direct and the
reverberation and the RT have become
half as long.
Additionally, in spite of the shorter RT,
the D/R has decreased from about -6 in
the large BSH model, to about -8.5 in
the half-size model.
LOC
=0.5
This is because the reverberation
builds-up quicker and stronger in the
smaller hall.
The direct sound, which was distinct in more than 50% of the seats in the large hall
will be audible in fewer than 30% of the seats in the small hall.
If the client insists on increasing the RT by reducing absorption, the D/R will be
further reduced, unless the hall shape is changed to increase the cubic volume.
The client and the architects expect the new hall to sound like BSH – but they, and
the audience, will be disappointed. As Leo Beranek said about the Berlin
Philharmonie: “They can always sell the bad seats to tourists.”
Great Small Halls Exist!
Jordan Hall at New England
Conservatory has 1200 seats, an
RT of 1.3s fully occupied. The shape
is half-octagonal, with a high ceiling.
The audience surrounds the stage,
with a single high balcony. The
average seating distance is much
shorter than a shoebox hall,
increasing the direct sound.
The high internal volume allows a
longer RT with low reverberant level.
The sound in nearly every seat is clear and direct, with a marvelous surrounding
reverberation. But the stage house is deep and reverberant. Small groups always
play far forward.
Although the hall is renowned as a chamber music hall, it is also good for small
orchestras and choral performances. It was built around 1905.
The hall is in constant use – with concerts nearly every night, (and many afternoons.)
Williams Hall, NEC
• Williams hall, in the same building, has ~350 seats in a square plan
with a high ceiling.
• The sound from a piano sound is clear and reverberant in most, if
not all, seats.
(The audience usually sits where the
orchestra is rehearsing in this picture.)
The square plan keeps the average
seating distance low.
The high ceiling and high single balcony
provides a long RT without a high
reverberant level.
The absorbent stage eliminates strong
reflections from the back wall. By
absorbing at least half the backward
energy from the musicians, the stage
increases the d/r.
Note the coffered ceiling – similar to
BSH.
Hard learned lessons
•
Where clarity is a problem in small halls, acousticians usually recommend adding
early reflections – through a stage shell, side reflectors, etc.
–
•
We tried this in a small hall by placing plywood panels behind the piano. The sound became
louder and less clear. Just the opposite of what was needed.
These measures reduce the gap between the direct sound and the reflected energy
and decrease LOC.
–
–
They increase loudness – which is usually already too high, while increasing the sense of
distance to the performers.
A better solution is to add absorption, or perhaps some means of deflecting the earliest
reflections to the ceiling, or into the front of the audience where they can be absorbed.
•
–
–
•
Re-direction tricks of this nature do not work well in small halls, as the second and third order
reflections they create will arrive within the 100ms window that determines LOC.
Small halls have strong direct sound and too many early reflections The early reflections
also come too quickly. Adding more reflections is exactly the wrong thing to do.
Adding absorption will improve clarity but reduce the late reverberant level and the RT.
Electronics, or more cubic volume, can restore the longer RT without decreasing the D/R.
In practice, not everyone is aware of, or appreciates, engagement. It is mostly a
subconscious perception. Reverberation or resonance is immediately apparent to
everyone – which is why it has become so over-emphasized in hall design.
–
–
Adding absorption may not be appreciated by everyone unless the decrease in late
reverberation can be compensated.
Such compensation can be surprisingly easy. Adding a few tenths of a second to the late
reverberation time of a small hall can be accomplished electronically with very few
loudspeakers. The result can be beautiful and completely transparent.
In the best halls the reverberant level is lower than
would be expected from classical acoustics
• D/R is frequency dependent in halls, and frequencies above 700Hz
are particularly important for engagement.
– Surface features can be used to decrease the reflected energy level in
the rear of the hall at higher frequencies.
• In addition, the distribution of absorption in a hall significantly alters
the distribution of the reflected energy.
– In a good hall absorption is highly non-uniform. A high ceiling with a lot
of reflecting surfaces above the audience can increase RT without
increasing the reflected energy level near the audience. The
reverberation created tends to stay up near the ceiling.
– This helps to keep the D/R above ~700Hz constant over a large number
of seats.
– Current modeling techniques may not properly calculate these effects.
• Old fashioned light models might work better…
Hall Shapes and direct-sound perception threshold
as a function of size
Above
threshold
Near
threshold
Below
threshold
A large hall like Boston
has many seats above
threshold, and many
that are near threshold
If this hall is reduced in
size while preserving
the shape, many seats
are below threshold
It is better to use a design
that reduces the average
seating distance, using a
high ceiling to increase
volume.
Boston is blessed with two 1200 seat halls with the third shape, Jordan Hall at
New England Conservatory, and Sanders Theater at Harvard. The sound for
chamber music and small orchestras is fantastic. RT ~ 1.4 to 1.5 seconds.
Clarity is very high – you can hear every note – and envelopment is good.
Retro reflectors above 1000Hz
Boston, Amsterdam, and
Vienna all have side-wall and
ceiling elements that reflect
frequencies above 1000Hz
back to the stage and to the
audience close to the stage.
This sound is absorbed –
reducing the reverberant level
in the rear of the hall without
changing the RT.
Another classic example is the
orchestra shell at the
Tanglewood Music Festival
Shed, designed by Russell
Johnson and Leo Beranek.
Many modern halls lack these
useful features!!!
High frequency retro reflectors
Rectangular wall features scatter in three
dimensions – visualize these with the
underside of the first and second
balconies.
High frequencies are reflected back to the
stage and to the audience in the front of
the hall.
The direct sound is strong there. These
reflections are not easily audible, but they
contribute to orchestral blend.
But this energy is absorbed, and thus
REMOVED from the late reverberation –
which improves clarity for seats in the back
of the hall.
Examples: Amsterdam, Boston, Vienna
High frequency overhead filters
A canopy made of partly open surfaces
becomes a high frequency filter.
Low frequencies pass through, exciting the full
volume of the hall.
High frequencies are reflected down into the
orchestra and the audience, where they are
absorbed.
Examples: Tanglewood Music Shed, Davies
Hall San Francisco
In my experience (and Beranek’s) these
panels improve Tanglewood enormously.
They reduce the HF reverberant level in the
back of the hall, improving clarity. The sound
is amazingly good, in spite of RT ~ 3s.
In Davies Hall the panels make the sound in the dress circle and balcony
both clear and reverberant at the same time. Very fine…
(But the sound in the stalls can be loud and harsh.)
The necessity of occupied measurements
• The effects of frequency dependent reflecting elements depends on
the presence of absorption on the stage and the front of the
audience.
• Measuring the halls without absorption in these areas will not detect
these vital effects.
• In addition, engagement is highly dependent on the D/R ratio – and
this is also not correctly measured in an unoccupied hall.
• Thus measurement of localization and engagement requires that
both hall and stage be occupied!
Binaural Measures
The author has been recording
performances binaurally for
years.
Current technology uses probe
microphones at the eardrums.
We can use these recordings to
make objective measurements
of halls and operas.
The hearing model described in
part one can be used to
measure the phase coherence
in these recordings.
Some demos of eardrum
recordings
•
These recordings have been equalized for loudspeaker reproduction. You
may be able to judge clarity and intelligibility over near-field loudspeakers.
– Accurate headphone reproduction requires headphone equalization
– A method of equalizing headphones through equal loudness measurements is
described in another paper on the author’s web-site.
– In general large circumaural headphones do not work well even when equalized.
On-ear phones, such as the Sennheiser 250 or 100, can work better.
•
opera balcony 2, seat 11
– Moderate intelligibility, reverberant sound.
– OK for non-Italian speakers with subtitles
•
opera balcony 3, seat 12
– Poor intelligibility, very reverberant
•
opera standing room
– Deep under balcony 2 – good intelligibility
– This was preferred by Italian speakers
•
A concert hall – row 8 (quite close)
– Very good sound. Not so good further back.
•
Conclusions
Performance venues should maximize engagement over a wide range of seats, while at
the same time providing adequate late reverberation. To achieve this goal the direct
sound must be perceived by the brain as distinct from the reflected energy – and this
includes early reflections from all directions.
–
–
•
The perception of reverberance and envelopment also depends on the audible presence
of direct sound.
–
•
–
–
–
The D/R ratio must increase as hall size is reduced if clarity, localization, and the sense of
envelopment is to be maintained.
D/R and engagement can be increased by decreasing the average seating distance, decreasing
the reverberation time, increasing the hall volume, or by careful use of rectangular diffusing
elements.
This is particularly true in opera houses and halls designed for chamber music.
A 1.8 second reverberation time is NOT necessarily ideal in a 1000 seat hall!!! Remember that
changes in reverberant LEVEL (D/R) and initial time delay are more audible than changes in RT.
To maintain clarity, low sonic distance, azimuth detection and envelopment in a small
hall (and many large halls) it is desirable to reduce the average seating distance, and
widely diffuse or absorb the earliest reflections, whether lateral or not.
–
•
Hall sound can often be improved by frequency dependent reflecting elements, or by adding
absorption to the stage rear wall.
The optimum value for the D/R ratio depends on the hall size –
–
•
In the presence of adequate late reverberation direct sound increases envelopment and reverberation loudness.
The audibility of direct sound depends on the D/R ratio above 700Hz, and the time delay
of reflections in the first 100ms.
–
•
Engagement is a spatial property that is sensitive to medial reflections. It can be heard with only
one ear (and measured with one microphone).
But it is essential to measure with both hall and stage OCCUPIED!
The best small halls do this already.
Current hall measurements ignore both the D/R and the time delay between direct sound
and reverberation. This talk introduces methods to overcome this lack.
Appendix
• Slides cut from the original presentation. They may – or may not, be
helpful to understanding some of the concepts.
Number of taps
• In our model the number of taps for each frequency is simply the
number of taps that fits in the delay line given the period we wish to
measure.
– If W is the length of the delay line in seconds, and f is the frequency in
Hz, then N, the number of taps is:
•
N(f) = W*f
• In our model the output of each summing neuron is divided by N, so
all frequencies have roughly the same amplitude.
• In a comb filter of this type, the frequency sharpness of each
summing node (full width at half maximum - fwhm) is just the
frequency divided by N.
•
fwhm = f/N
• In practice the sharpness is higher, as the input to the filter is not
sinusoidal, but sharply peaked at the frequency of the fundamental.
• Our model achieves a frequency discrimination (+-3dB) of 1%.
Detection of harmonics and sub-harmonics
• Comb filters are sensitive to harmonics of the tap period.
– The full width at half maximum is sharper for the harmonic than
for the fundamental.
– Thus a modulation that contains harmonics will be more
accurately detected.
• A tap sequence intended to detect 100Hz will also detect
200Hz and 300Hz.
– This property produces what appear to be sub-harmonics of the
input modulations, since a 200Hz modulation will also excite the
100Hz tap sequence.
• This artifact has interesting consequences for harmony,
as the pitch pattern produced by both major and minor
triads has a strong output at the root frequency
– Regardless of the inversion of the notes in the triad.
Minor triad in two inversions
Solid line: output of the pitch
detector with a minor triad,
200Hz, 235Hz, and 300Hz.
Dotted line: The same triad
with the fifth lowered by one
octave.
200Hz, 235Hz, and 150Hz.
Notice the acuity of the pitch
detection is better than 1%,
or 1/6th of a semitione.
Each peak represents a
separate neural data stream,
which can be further
analyzed for timbre and
azimuth.
Parameters
The model contains several parameters:
• 1. The choice of a log-linear model for the hair cells, not
a fully logarithmic detector. (the choice is clear…)
–
1600,2000Hz bands after hair cell: Source - Logarithmic
Log-linear
• 2. The choice of 10ms as the response time of the
logarithmic adaptation.
– This causes the onsets of sounds to have a higher amplitude
than the body of the sound, which is probably an advantage
• 3. The choice of a 100ms window for the frequency filter
– This choice is consistent with earlier work on localization.
• 4. The use of equal weighting for all the taps
– Chosen for simplicity. Further work is needed. But equally
weighted taps work rather well…
Why pitch is robust in the presence of
reverberation
• Although pitch acuity is reduced by reverberation,
reverberation is not capable of completely scrambling
the phases of the harmonics.
• Some 1/3 octave bands will still contain significant
modulation – but which have modulation, and which do
not, is random, and varies from note to note.
• If we are relying on the amplitude of the modulations in
each band to determine timbre – then timbre will be
scrambled.
• In fact the major perception of a distant source is that the
timbre has become muddy.
What is “Auditory Engagement”
•
“Engagement” is the perception that you are not just watching a scene from
distance, but present in the middle of it.
– Thus lack of distance is a critical component of presence.
•
Auditory engagement is the perception that you are acoustically close to the
sound sources.
– Distance is perceived directly through harmonic coherence – but experiments to
directly measure it with subjects are difficult. However it correlates both with the
ability to localize sound sources, and the perception presence, or musical clarity.
– To perceive presence you must be able to localize sound sources nearly all the
time,
– and be able to distinguish them from one another nearly all the time.
•
Clear localization and the ability to hear most of the notes are key
components of audience engagement.
– Although particularly important in drama and opera, it should be (and often is not)
a part of the emotional experience of music.
– Being able to hear all the notes and localize the players draws the audience into
the performance. They don’t just watch it.
•
This view of clarity is different from the one that equates clarity with
intelligibility. Perhaps we need a new word for it.
Experiment for threshold of Azimuth
Detection in halls
A model is constructed with a
source position on the left, and
another source on the right
Source signal alternates between
the left and a right position.
When the d/r is less than about
minus 13dB both sources are
perceived in the middle.
Subject varies the d/r, and reports
the value of d/r that separates the
two sources by half the actual
angle.
This is the threshold value for
azimuth detection for this model
(Above this threshold the subject also reports a decrease in subjective distance)
Threshold for azimuth detection as a
function of frequency and initial delay
As the time gap between the direct
sound and the reverberation
increases, the threshold for azimuth
detection goes down. (the d/r scale
on this old slide is arbitrary)
As the time gap between notes increases
(allowing reverberation to decay) the
threshold goes down.
To duplicate the actual perception in small
halls I need a 50ms gap between notes.
An important caveat!
• All these thresholds were measured without visual cues
• The author has found that in a concert (with occasional visual input)
instruments (such as a string quartet) are perceived as clearly
localized and spread.
• When I record the sound with probes at my own eardrums, and play it
back through calibrated earphones the sound seems highly accurate,
but localization often disappears!
– Without visual cues when the d/r is below threshold the individual
instruments are localized and spread when they play solo, but collapse to
the center when they play together.
– My brain will not allow me to detect this collapse when I am in the concert
hall – even if I close my eyes most of the time!
– With eyes closed it is more difficult to separate the sounds of the
individuals, such as the second violin and the viola. This difficulty persists
in the binaural recording.
Localization
• For this paper we assume sound sources are localized by the direct
sound.
– In some cases localization is aided by early reflections – but these vary
strongly from seat to seat, and are too complex to consider here.
• For localization to be successful the direct sound must be perceived.
– Prompt strong reflections can – and do – mask the direct sound.
• Let’s propose that the brain detects the loudness of – and the
presence of – sounds by integrating nerve firings over a period of
time.
– If the integrated nerve firings from the direct sound exceed the
integrated nerve firings from the reflections inside this time window, the
direct sound will be perceived – and localized.
• We can calculate the threshold of perception by double integrating
the impulse response over a fixed time window.
The ear perceives notes – not the impulse
response itself.
•
Here is a graph of the ipselateral binaural impulse response from spatially diffuse
exponentially decaying white noise with an onset time of 5ms and an RT of 1 second.
This is NOT a note, and NOT what the ear hears!
D/R = -10dB
RT = 2s:
C80 = 3.5dB
C50 = 2.2dB
IACC80 = .24
RT = 1s:
C80 = 6.4dB
C50 = 4.1dB
IACC80 = .20
•
To visualize what the ear hears, we must convolve this with a sound.
–
Let’s use a 200ms constant level as an example.
•
The nerve firings from the direct component of this note have a constant rate for the
duration of the sound.
•
The nerve firings from the reverberant component steadily build up until the note ceases
and then slowly stop as the sound decays.
Direct and reverberation for d/r = -10dB, and RT = 1s
The blue line shows the rate of nerve firing rate for a constant direct sound 10dB less than
the total reverberation energy. The red line shows the rate of nerve firings for the
reverberation, which builds up for the duration of the note. The black line shows a time
window (100ms) over which to integrate the two rates. In this example the area in light
blue is larger than the area in pink, so the direct sound is inaudible.
Direct and build-up RT = 2s
If we hold the d/r constant, when the reverberation time is two seconds it takes
longer for the reverberation to build up, so the light blue area decreases, while the
pink area stays constant. This makes the direct sound more audible. In a large
hall the time delay between the direct sound and the reverberation also increases,
further reducing the area in light blue. The direct sound would be even more
audible.
Equation for Localizability – 700 to 4000Hz
•
We can use this simple model to derive an equation that expresses the
ease of perceiving the direction of direct sound as a decibel value. p(t) is the
sound pressure of the ipselateral channel of a binaural impulse response.
With the previous simple assumptions, we propose the threshold for
detection would be 0dB, and clear localization would occur at a localizability
value of +3dB.
•
Where D is the window width (~ 0.1s), and S is a scale factor:
S is the zero nerve firing line in the previous
two slides. It is 20dB below the maximum
loudness. POS means ignore the negative
values for the sum of S and the cumulative
log pressure.

 p(t ) dt
S  20  10* log
2
.005
•
Localizability (LOC) in dB =
.005
S  1.5  10* log 
0
•
p(t ) dt  (1 / D) * 
2
D .005
0

POS(S  10* log 
.005
p(t ) 2 dt)d
The scale factor S and the window width D interact to set the slope of the
threshold as a function of added time delay. The values I have chosen
(100ms and -20dB) fit my personal data. The extra factor of +1.5dB is
added to match my personal thresholds.
Some explanation of the equation
•
•
The equation as written in the previous slide simply calculates the ratios of
the pink and blue areas shown in the previous pictures.
The first integral on the left in LOC is the “pink” area – the sum of the nerve
firings for the direct sound. This area is the product of the normalized
sound pressure times the length of the window D.
– However here we have divided through by D – so this factor is not shown.
•
The next two integrals represent the total nerve firings for the reverberation
– the “blue” area.
–
•
•
•
Since we have divided by D, a factor of 1/D is included at the beginning.
The second of the two integrals is the physical sum of the sound pressure
that would exist if the impulse response was convolved with a steady
excitation. The first integral finds the area under this curve. In the second
integral we have excluded the direct sound – assuming this will be in the
first 5 milliseconds.
The limits of the integrals have been adjusted to account for this exclusion.
Thus the second integral goes from .005 seconds to the end, and the first
integral is from zero to the window width minus .005.
I have included the -1.5dB adjustment for my personal thresholds.
Matlab code for LOC
% load in a .wav file containing a binaural impulse
response – filter it and truncate the beginning
upper_scale =20; % 20dB range for firings
% proposed box length
box_length = round(100*sr/1000); % try 100ms
early_time = round(5*sr/1000);
% ir_left is an ipselateral binaural impulse response,
%truncated to start at zero and filtered to 1000-4000Hz.
% early_time is 5ms in samples, D is 100ms in samples.
% here starts the equation on the slide:
D = box_length;
%the window width
ir_left = data1; % the binaural IR
ir_right = data2;
clear data1 data2 % filter the Irs
wb = [2*1000/sr 2*4000/sr];
[b a] = ellip(3,2,30,wb);
ir_left = filter(b,a,ir_left);
ir_right = filter(b,a,ir_right);
clear data1 data2
wb = [2*1000/sr 2*4000/sr];
[b a] = ellip(3,2,30,wb);
ir_left = filter(b,a,ir_left);
ir_right = filter(b,a,ir_right);
for il = 1:0.1*sr
if abs(ir_left(il)) > 500
break
end
if abs(ir_right(il)) > 500
break
end
end
ir_left(1:il) = [];
ir_right(1:il) = [];
S = 20-10*log10(sum(ir_left.^2));
early = 10*log10(sum(ir_left(1:early_time).^2));
% first integral is a cumsum representing the build up in
%energy when the IR is excited by a steady tone:
ln = length(ir_left);
log_rvb = 10*log10(cumsum(ir_left(early_time:ln).^2));
% look at positive values of S+log_rvb only
for ix = 1:ln-early_time
if S+log_rvb(ix) < 0
log_rvb(ix) = -S;
end
end
LOC = S-1.5+early -(1/D)*sum(S+log_rvb(1:D-early_time))
Use of the localization equation
• Just as RT or C80, LOC uses a measured impulse response as an
input, with the direct sound starting at time zero. This is the only data a
user needs to supply.
– The measure is calibrated for a front facing binaural impulse response.
• An omnidirectional impulse response will give lower values of LOC for the same
seat position, due to the lack of head shadowing.
• The localization equation appears more complex than most current
measures for room acoustics, but it has a simple, physiologically based
interpretation.
– It is the ratio in dB of the number of nerve firings received by the brain from
the direct sound in a 100ms window, divided by the number of nerve firings
received from all reflections in the same time period.
– It contains three experimentally based parameters: the window width D, the
dynamic range of the nerve channels S, and the time window for separating
direct sound from reflections (5ms). These parameters are not intended to
be adjustable without further experimental work.
– Matlab code for calculating LOC is simple, as can be seen above.
Interpretation of LOC
•
•
LOC was developed and verified as a method for predicting when a sound
will be accurately localized when the direct sound is much lower in total
energy than the sum of all reflections.
Like C80, IACC80, and similar measures, LOC is based on a time window
that begins with the onset of the direct sound.
– In practice, syllables or notes that will be affected by any of these measures will
depend on the rise time (onset time) of the sound.
– If the sound starts gradually the precise moment of onset becomes
indeterminate, and separating direct sound from reflections becomes impossible.
– Thus LOC – and other such measures – are accurately predictive only for signals
with sharp onsets.
– Additionally, if the direct sound from a note or syllable is masked by
reverberation from a previous sound, the direct sound will not be audible.
•
LOC predicts the audibility of the direct sound for a syllable or note with a
rapid rise-time when there is sufficient freedom from masking from previous
sounds.
– Although musical signals often do not meet these criteria, in practice there are
enough occasions that do meet the criteria that the LOC equation is useful.
•
Remember that for the purposes of this talk Localization is only a proxy for
the main goal – predicting when the direct sound is sufficiently audible to
produce engagement.
– Preliminary results suggest LOC achieves this goal.
Localization Equation Setup
•
The Localization Equation was developed and tested using binaural impulse
response generated using the author’s own HRTFs.
– The source position was 15 degrees to the left (and right) of center. Only the
ipselateral channel was analyzed.
– Male speech alternated from left to right with a time gap of 400ms, to allow for
complete decay of the reverberation between each word.
– The reverberation was generated using an independent decaying noise signal
convolved with each of 54 HRTFs spaced equally around the listening position.
– The HRTFs were equalized so that the azimuth zero elevation zero HRTF was
flat from 40Hz to about 4kHz. The elevation notch at 7.8kHz was not equalized
away, but was left in place.
– Playback was done through headphones equalized to match a loudspeaker
placed in front of the listener – again not equalizing the 7.8kHz notch from the
listener’s frontal HRTF of the loudspeaker.
• Because my data show that the perception of both localization and
near/far is mostly a high frequency phenomenon, the impulse
response was bandpass filtered between 700Hz and 4000Hz before
being analyzed for localization.
– If a measured binaural impulse response is used as an input, care should be
taken to insure the dummy head is equalized as described above.
– Because of the importance of upward masking in localization, if the low
frequencies in the room signal are significantly stronger than those in the
frequency range from 700 to 4000Hz, localization is likely to be poorer than the
equation would predict.
Comments on LOC
– LOC is based on the LOG of the build-up of reverberant energy.
• This follows directly from the physiological model.
• Current measures integrate the sound energy rather than the log of sound energy. But our
physiology works differently. One of the consequences is that reflections that arrive early
have more influence than reflections that arrive later.
– As energy builds up additional reflections are not counted as strongly.
• Reflections later than 100ms are ignored in calculating LOC.
– This is very different from C80 or C50, which count the earliest reflections a part of
the direct sound, and compare the energy sum to the energy sum of all the later
reverberation.
• In a small hall most of the energy arrives before 80ms regardless of the relative strength
of the direct sound, so C80 and C50 are usually high.
• But small halls can have high C80 or C50, poor localization, and a lack of clarity.
– LOC depends strongly on the delay between the direct sound and the build-up of
the reverberation.
• late reverberation does not impair localization of short notes.
• The principle difference between the localizability in small halls and large halls is the rate
at which reflected energy builds up after the start of a note.
– LOC is NOT related to EDT – even if Jordan’s original definition of EDT is used.
• EDT is relatively independent of the initial time delay
• When D/R < -10dB, EDT and RT are the same, as there is insufficient direct sound to be
detected in a reverse integrated impulse response.
– LOC correlates with IACC80 – but IACC is not sensitive to medial reflections.
• IACC is sensitive to the sum of reflected energy – not the log of energy, and thus is
insensitive to when the reflections arrive
Tests with speech
A speech signal was convolved with a pair of binaural impulse responses,
such that the sound appears to come from +-15 degrees from the front.
Then a fully spatially diffuse reverberation was added, in such a way as the
D/R, the RT, and the time delay before the reverberation onset could be
varied.
Broadband Speech Data
Blue – experimental thresholds for the alternating speech with a 1 second reverb
time. Red – the threshold predicted by the localization equation. Black –
experimental thresholds for RT = 2seconds. Cyan – thresholds predicted by the
localization equation.
Threshold Data from Other Subjects – 1s RT
Blue – new data using
absence of any
localization as a
criterion for threshold.
Red – the author’s
previous data based
on a half-angle
criterion.
• Seven subjects participated in a threshold experiment at Kyushu
University.
– In these experiments the threshold was defined by the extinction of
localization , not by the reduction of angle by a factor of two.
– Consequently the thresholds are lower than they were in my previous
experiment, and they have more variation.
• However, the data is consistent to within 3dB
Analysis Add reverb at 2s RT -10dB
D/R
Two second RT at -10dB D/R does not seriously interfere with pitch.
Add reverb at 1s RT -10dB D/r
RT = 1s is much more damaging to harmonic coherence in the 100ms window.
An existing small hall – pictures
Note the highly reflective stage and side
walls, deeply coffered ceiling, and relatively
low internal volume per seat.
The sound in many seats is muddy. Adding
reflections or decreasing absorption only
increases the muddiness.
Hall data
• The pictures show a recital hall of 65000 cubic feet (1840 cubic
meters). Designed for 350 seats, it has currently 300 seats, giving a
volume/seat of 6 cubic meters. There is 1400 square feet of carpet
under the seats on the floor.
• Reverberation Time (RT) unoccupied is 1.1 seconds from 1000Hz to
63Hz. C80, dominated by the reverberation time, is ~+5dB
everywhere.
• The parallel side walls of the stage provide little diffusion.
• The hall is generally liked by the audience and players, although
there are reports of loudness and balance problems on stage.
• Musicians desire more resonance and greater clarity in the middle of
the hall.
Experiments with absorption and
acoustic enhancement
•
Measurements and experiments involving various combinations of
fiberglass panels and electronic reverberation enhancement were
conducted in January 2009.
– Measurements were made with three loudspeakers, three dummy heads, and a
Soundfield microphone.
– All musical performances were recorded with the same microphones, and with
an array of close microphones on stage.
•
About 30 musicians participated, including faculty, staff, students from all
three divisions, and musicians from the wider community.
•
The goal was to improve the instrumental balance on stage, reduce excess
stage loudness, and to increase resonance and the ability to hear individual
instruments throughout the hall.
•
With both panels and enhancement in place comments from the participants
were favorable. Players and singers found balancing with piano was easier,
and the middle registers of the piano were more easily heard both by the
musicians and in the hall.
The absorptive curtains at the rear of the stage could be rapidly withdrawn. The blankets that
simulated audience could be removed in 5 minutes, along with the panels on stage. This
allowed prompt A/B comparisons. Some of the 25 LARES enhancement speakers are visible
Results from the experiments
•
The experiments in January showed that adding fiberglass panels around
the stage increased clarity and the ability to localize instruments in the hall,
raising the measured value of LOC from an average of minus 1.5dB to +3dB
or more.
•
Localization and clarity in the balcony were additionally improved by adding
panels to the upper audience right side wall, which eliminated the strong
lateral reflection from that surface.
– The lower surface of this wall was already absorptive.
•
The electronic enhancement successfully compensated for the loss of
resonance due to the panels. Without the enhancement the perceived
resonance was reduced.
•
In a subsequent experiment with a violin-piano combination and no
enhancement we found that just 12 fiberglass panels each 2’x6’x2” around
the bottom of the stage noticeably improved the clarity on the floor of the
hall, and also improved the balance for the players on stage. For this music
the reduced resonance was not a problem.
– These panels absorbed the first-order reflection from the back of the stage,
which has the highest level and the shortest time delay. Absorbing this reflection
contributed strongly to increasing LOC.
Usefulness of the measure LOC
• LOC informs us that the primary contribution to difficulty in
localization are the first strong reflections, regardless of the direction
they come from.
• We initially thought that since the floor of the hall is not a significant
source of these reflections, it is would be likely that removing the
carpet under the seats would raise the RT without decreasing LOC
significantly.
– However LOC is also sensitive to reverberation which arrives before
100ms, and this would be increased by removing the carpet.
– A few later experiments suggest that removing the carpet will increase
the reverberant level sufficiently to eliminate the improvement in LOC
provided by the absorption on stage.
• The existence of a LOC as a physical measure can help to answer
these questions in advance – or at least suggest that an experiment
is needed before drastic alterations are undertaken.
Small shoebox halls can be OK
• If the client insists on a shoebox it can work by building a
large hall and installing a small number of seats.
– I was just in such a small hall in Helsinki, and at least half the
seats were OK.
• But this is not the ideal solution.
– With a different shape nearly all the seats could have been OK –
and it might have been less expensive.
Light models
I ran across these pictures while
cleaning out my office. The top
one is a too-simple model of the
Philadelphia Academy of Music.
The bottom is intended to be
BSH, but with a single balcony.
I abandoned light modeling
because it does NOT provide any
information about the time delay
gap – nor information about the
effects of note length on D/R.
But it DOES provide information
about the total reverberant energy
compared to the direct. And very
complex hall shapes can be
quickly modeled.
A few slides from earlier measures of azimuth from
recordings, using running IACC in a 10ms window.
They are followed by an earlier measure of
harmonic coherence.
Localization
The figure shows the number of
times per second that a solo violin
can be localized from row 4 of a
small shoebox hall (~500 seats)
near Helsinki.
It also shows the perceived
azimuth of the violin
As can be seen, the localization –
achieved at the onsets of notes –
is quite good, and the azimuth,
~10 degrees to the left of center,
is accurate.
Localization – surface1
Here we plot the same
data for the violin as a
function of (inverse)
azimuth, and the third
octave frequency band.
As can be seen, for this
instrument the principle
localization components
come at about 1300Hz.
Interestingly, Human ability
to detect azimuth, as
shown in the threshold
data, may be maximum at
this frequency.
Localization, Surface 2
Here we plot 1/(1-IACC) as
a function of time and third
octave band.
Note that the IACC peaks at
the onset of notes can have
quite high values for a brief
time.
This happens when there is
sufficient delay between the
direct and the reverberation,
and sufficient D/R.
Localization – a poor seat
Here is a similar diagram for
a solo violin in row 11 of the
same hall. The sound here
is unclear, and the
localization of the violin is
poor.
As can be seen, the number
of localizations per second is
low (in this case the value
really depends on the setting
of the threshold in the
software).
Perhaps more tellingly, the
azimuth detected seems
random.
This is really just noise, and
is perceived as such.
Measures based on harmonic coherence
•
In the absence of reflections the formant frequencies above 1000Hz are
amplitude modulated by the phase coherence of the upper harmonics. This
modulation is easily heard, creating the perception of “roughness” (Zwicker).
– Reflections randomize the phase of these harmonics.
•
•
The result is highly audible, and is a primary cue for the distance of an
actor, singer, or soloist.
This effect can be measured with live recordings, and is sensitive both to
medial and lateral reflections.
This graph shows the frequency and
amplitude of the amplitude modulation of a
voice fundamental in the 2kHz 1/3 octave
band. The vertical axis shows the effective
D/R ratio at the beginning of two notes from
an opera singer in Oslo to the front of the
third balcony (fully occupied.) The sound
there is often muddy, but the fundamental
pitch of this singer came through strongly
at the beginning of these two notes. He
seemed to be speaking directly to me, and
I liked it.
Another singer
From the same seat the king (in Verdi’s
Don Carlos) was not able to reach the
third balcony with the same strength.
Like the localization graph shown in a
previous slide, this graph seems to be
mostly noise.
The fundamental pitches are not well
defined. The singer seemed muddy
and far away.
His aria can be heart-rending – but
here it was somewhat muted by the
acoustics. We were watching the king
feel powerless and forlorn. But we
were not engaged.