MODULE 1: IMPULSE RESPONSE MEASUREMENTS

Guillermo González Jiménez, Eva Frossard, Mingyang Gao, Helene Vargas Becerra

Group 3

ABSTRACT 2. ROOM IMPULSE RESPONSE MEASUREMENTS

Experimental setup and measurement framework

In this report, we conducted measurements and data analy-
sis on the impulse response (IR) of a reverberant chamber A sketch of the measurement setup and a picture of the rever-
in DTU. This specifically includes a description of the ex- berant chamber where the experiment took place are shown
perimental setup and measurement framework, as well as in the figure 1.
the measurement principles and objectives of the energy time
curve (ETC). In the analysis section for the IR, we expounded
on the formula and data processing for the energy decay curve
(EDC), the fitting methods used for estimating reverberation
time models, and the final results. We also calculated the
direct-to-reverberant energy ratio (DRR) and plotted the en-
ergy decay relief (EDR). The last section involves the creation
of binaural room impulse responses (BRIR) signals, using the
measured IR data to convolve with the signal for effect en-
hancement.

Index Terms— Impulse response, Energy time curve, En-

ergy decay curve, Reverberation time, Direct-to-reverberant
energy ratio, Binaural room impulse response

(a) Overview of the measurement setup

1. INTRODUCTION

Many acoustical systems can be modeled as linear time-

invariant (LTI) systems. Under appropriate constraints, the
measurement and subsequent analysis of their impulse re-
sponses (IRs) are effective means of characterization. Under-
standing the behavior and characteristics of these acoustical
systems enables us to make informed decisions in a wide
range of applications: room acoustics, loudspeaker design,
audio signal processing, etc ...
During this practice, the studied acoustical system is a
highly reverberant room. Starting with the measurement
setup and procedures, we will progress to analyzing the data,
including the Energy Decay Curve to characterize precisely
the decay of sound energy and to estimate the reverberation
(b) Reverberation chamber in Building 355 at DTU
time, the Direct-to-Reverberant Energy Ratio to provide a
quantitative measure of the direct sound-to-reverberant sound Fig. 1: Measurement setup used for the experiment
ratio in the room, and finally the Energy Decay Relief to
enable frequency-specific energy decay analysis. Finally, we
will also convolve our measurements with an anechoic signal.
 The following equipment was used: Running the measurement

• two microphones To estimate the noise floor of the reverberation room, we

played the excitation signal and plotted its corresponding
• a head and torso simulator (HATS) energy time curve (ETC) :

ETClog (t) = 10 log10 (h2 (t) (1)

• an amplifier
As first measurement, we played excitation signal with a
• a sound-card duration of 1 second followed by 1 second of silence. The
resultant ETC is shown in figure 2
• a loudspeaker

• a laptop computer running MATLAB interfaces

The two microphones were aligned with the loudspeaker

at 1 and 2 m distance, and the HATS was facing the sound
source.
The MATLAB measurement framework used to measure
the impulse response in the reverberant room was composed
of the following files:

• Measure IR.m : Template for running a measure-

ment

• genMeasSig.m : Generates a measurement signal

using the genChirp() function

• genChirp.m : Creates a chirp signal Fig. 2: Energy time curve of measurement 1

• fade.m : Applies a raised cosine window to the input The initial and highest spike in the ETC corresponds to the
signal for fade in and fade out sound that travels directly from the loudspeaker. The initial
pikes of the two microphones are not recorded at the exact
• padZeros.m : Zero-pad the input signal same time. The microphone located one meter away from the
sound source (channel 1) is the first one to show that pike.
• getInverse.m : Creates an inverse filter in the fre- After the direct sound, we can notice smaller peaks on the
quency domain ETC, corresponding to early reflections. After a few seconds,
the peaks become less distinct. Those ones correspond to the
• getIR.m : Estimates the impulse response h by ap- late reflections and reverberation. The intensity of the peaks
plying the inverse filter keeps decreasing until the curve reaches the level of the noise
floor.
Considering the reverberation time and the noise floor In the figure 2 we can see that the noise floor after the
of the room, we decided to use an initial length of 6s for sweep signal is quite high. This is because all the noise of the
the sweep signal. Using this initial length, the signal is long sweep does not have time to fully decay during the silence
enough to capture the full decay of the room’s reverberation, period.
which we estimated at 4s, it is also long enough so that the If the silence period is shorter than the time it takes to the
energy level is significantly above the noise floor and we signal’s frequencies to decay, then the signal’s spectrogram
can easily distinguish the impulse response from the ambient after convolution processing will contain noise and the end of
noise. the signal will decrease abruptly, as shown in figure 3
For this experiment, we decided to use an upward sweep,
because low frequencies take more time to decay. That way,
the low frequencies can partially decay while mid and high
frequencies are generated. After the highest frequency is
reached, the silence period starts allowing us to obtain accu-
rate impulse response measurements.
 (a) STFT of channel 1

Fig. 4: Energy time curve of measurement 3

(b) STFT of channel 2

Fig. 3: STFTs of measurement 1

A very short sweep signal can also lead to issues with

STFT’s circular convolution at the edges of the time window, (a) STFT of channel 1
which can produce artifacts at the end of the signal.
To fix this, we decided to increase the silence period. The
values that worked best were a sweep of 2 seconds and a si-
lence of 6 seconds. With these parameters we got the ETC of
Fig. 4, which has a lower noise floor, because the long silence
period allows the noise of the sweep to completely dissipate.
The energy corresponding to the signal will be too low to ob-
tain a good resolution.
This improvement can also be seen in the spectogram of
the signal (Fig. 5), which seems smoother and presents minor
artifacts at the end of the time axis.
By looking at the ETC of the different measurements, we
can see that all the dynamic ranges (Table 1) are sufficient,
but the best dynamic ranges correspond to the measurements
3 and 4. We decide to save the set of result number 3, because (b) STFT of channel 2
it has a shorter measurement time.
Fig. 5: STFTs of measurement 3
 Measurement Sweep Silence Dynamic little subsequent silence. The truncation time before remov-
number [s] [s] range [s] ing the noise floor at the start of the signal is shown below in
1 1 1 70 dB table 2.
2 2 4 78 dB
3 2 6 85 dB Measurement Sweep Silence Truncation
4 6 6 95 dB number [s] [s] time [s]
5 4 6 80 dB 1 1 1 1.67
2 2 4 3.65
Table 1: Dynamic range estimation for each measurement 3 2 6 3.70
based on the sweep length and decay 4 6 6 4.85
5 4 6 4.21

3. ANALYSIS OF THE ROOM IMPULSE RESPONSE Table 2: Truncation times selected for each measurement
based on the sweep length and decay
This section explores the measurements made in the reverber-
ant room to see how energy decays. First the EDC (energy
The EDC for measurement 3 - 2s sweep, 6s silence - trun-
decay curve) provides a wide-band characterization of how
cated at 3.70 seconds is plotted below in figure 6.
the energy decays which is then used to estimate the reverber-
ation time of the room. The measurements are then split into
two parts, the direct energy and reverberant energy to find the
direct-to-reverbarent energy ratio. Finally, the EDR (energy
decay relief) will be plotted to illustrate how the energy of
individual frequencies decay over time.

3.1. Energy Decay Curve

The energy decay curve (EDC) is a method to measure the
reverberation time more precisely, without the necessity to
repeat the reverberation experiment to minimize the effect of
random fluctuations in the decay curves [1].
To perform this method, the IR recording must be time
reversed, squared and then integrated [1]. This serves as filter
for random fluctuation while keep the exponential shape of IR
[2]. This procedure is defined as:

Rt 2 !
0
h (τ )dτ
EDCdB (t) = 10 log10 1− R∞ 2 (2) Fig. 6: Energy decay curve of channel 1 and 2 for measure-
0
h (τ )dτ
ment 3, 2s sweep, 6s silence and truncation time 3.7 s.
Although EDC provides a better method to estimate rever-
beration time it presents limitations regarding the noise floor.
As noise floor is an inevitable factor, it limits the minimum 3.2. Estimating reverberation time
level of the IR, which causes a distortion in the desired expo-
Once the IR is truncated to avoid errors related to noise floor,
nential shape of the IR [2].
the next step to estimate the reverberation time is to measure
A solution to deal with this is to truncate the IR signal.
the EDC slope of the truncated signal IR down to a level about
Even though there are methods to estimate the truncation
5 dB above the noise floor [2]. The reason for this limit is that
time, modern approach leans towards a more intuitive estima-
this range provides a balance between capturing the relevant
tion. This approach states that the signal should be truncated
information and reducing the impact of noise floor effects.
at a point where the main IR slope intersects the noise floor,
The reverberation time was estimated by fitting a linear
called knee point of the curve. The reason behind this choice
model to two points on the EDC, one at -5 dB, to avoid early
is that for evaluation ranges above noise floor it has small
reflections, and another at -25 dB, to have a decibel range of
bias, moderate sensitivity and a fitting error less than about 2
20 dB and therefore estimate the reverberation time T20 .
dB [2].
Knowing this, before calculating the EDC, the measure-
ment is truncated to include only the sweep, decay and only
 Measurement Channel 1 Channel 2
number DRR DRR
1 -0.26 -7.52
2 -0.33 -7.55
3 -0.32 -7.55
4 -0.34 -7.53
5 -0.33 -7.72

Table 4: The direct-to-reverberant energy ratios for each mea-

surement and channel

R td 2 !
h (t)dt
DRRdB = 10 log10 R 0∞ (3)
td
h2 (t)dt
Equation 3 gives the DRR in dB where td is the time just
before the reverberant sound reaches the receiver. Thus the
numerator only includes direct sound whereas the denomi-
Fig. 7: Energy decay curve of channel 1 for measurement 3, nator also include reverberant sound. Thus a very negative
2s sweep, 6s silence and truncation time 3.70 s. Fitted with a DRRdB value will mean little direct energy and a lot rever-
linear model to estimate T20 . berant energy. If the DRRdB is close to 0 then there is almost
as much reverberant energy as direct.
td was selected to be the time just before the second peak
The linear model thus estimates the rate at which energy in the ETC. The first peak is the direct sound reaching the
decays in the room and the time it would take for energy to microphone from the speaker whereas the second peak is the
decay by 60 dB. The estimated reverberation time T60 was earliest reflection. The earliest reflection would in our case be
thus calculated for each measurement and channel and the from the floor.
results are found in the next table 3. Table 4 show the DRRdB values for each measurement.
The DRRdB of the microphone closest to the speaker is much
Meas. Sweep Silence Channel 1 Channel 2 closer to zero than the one further away, thus, there is more
number [s] [s] T60 [s] T60 [s] reverberant energy the further away for the speaker the micro-
1 1 1 3.287 3.437 phone is.
2 2 4 5.252 5.548
3 2 6 5.242 5.343 3.4. Energy decay relief
4 6 6 5.443 5.985
5 4 6 5.204 9.834 The energy decay relief (EDR) is a spectral extension of
the EDC. EDR provides a time-frequency representation of
Table 3: Estimated reverberation time of reverberant room the energy decay. This results in a 3D illustration of the
frequency-specific energy decay of the measured impulse
response [3].
The results found for measurement 1 (table 3) is disre- In order to calculate the EDR of the impulse response, the
garded since the silence was too short for the energy to fully next steps were followed:
decay and thus the results are unreliable. The result for chan-
nel 2 in measurement 5 (table 3) also disregarded as the mea- • Truncate the IR to the estimated knee point as with
surements of that channel where messed up due to a equip- EDC calculation.
ment malfunction.
• Apply the STFT to the truncated IR.
Taking the average of the remaining reverberation times,
the estimated reverberation time (T60 ) is 5.431 seconds • Calculate EDR applying reversed-time integration of
squared STFT and transform it to dB scale.
3.3. Direct-to-reverberant energy ratio • Normalize EDR to 0 dB and truncate it to -60 dB.
The direct-to-reverberant energy ratio (DRR) compares the The EDR for measurement 3 are shown in figure 8.
relative amount of direct and reverberant sound for a given By looking the EDR plots, it can be seen the energy dis-
receiver and source in a room. tribution among frequencies and its dissipation through time
 4. BINAURAL IMPULSE RESPONSES

In our last measurement record file, we measured the BRIRs

(Binaural Room Impulse Responses) in the room using the
parameters: Tsilence=6S, Tsweep=6s, fs=48kHz.
Task: Create some reverberant binarual stimuli by con-
volving your measured responses with an anechoic signal of
your choice.
We choose the ’cockroach king.wav’ file as our anechoic
signal and load it using the readAudio function, and then con-
volve the signal with measured impulse responses. In order to
visualize the results, we performed FFT on the original signal
and the processed signal. The left part of Fig. 9 is the original
signal, and the right part is the processed signal.

(a) Energy decay relief of measurement 3 channel 1

(b) Energy decay relief of measurement 3 channel 2 (a) STFT of anechoic signal before convolution
Fig. 8: EDR of measurement 3 (2s sweep, 6s silence)

of the reverberation room. As it would be expected due to the

use of an increasing exponential sweep, low frequencies had
more energy at the beginning of the measurement, as a result
of the sweep spending more time in lower frequencies than in
higher ones.
As time goes by, the energy of all frequencies decays at
different rates, being the lower frequencies the ones that take
longer to decay and the higher frequencies the ones that take
shorter. From this behavior it can be inferred that the room ab-
sorbs the higher frequencies, causing a shorter reverberation (b) STFT of anechoic signal after convolution
time for this frequency band, and reflects the low frequencies,
resulting in a longer reverberation time for low frequencies. Fig. 9: Comparison of spectrograms before and after convo-
One factor to consider when analyzing the measurements lution processing
is the noise that was present. Although the measurements
were made in a controlled environment, it should be men-
tioned that the port of the room was not completely closed, 5. CONCLUSION
so external noise may have been present during the measure-
ments. Nevertheless, when looking at the results it can be said The purpose of this report is to document our measurement
that this is negligible. process and analysis results for the impulse response (IR)
of the reverberant chamber, which is an LTI system. We
obtained specific characterization parameters, namely EDC,
DRR, and EDR, and applied the measured IR to a signal.

6. REFERENCES

[1] Schroeder M. R., “New method of measuring reverbera-

tion time,” Journal of the Acoustical Society of America,
vol. 37, pp. 409–412, 1965.
[2] Dennis R. Morgan, “A parametric error analysis of the
backward integration method for reverberation time esti-
mation,” The Journal of the Acoustical Society of Amer-
ica, vol. 101, no. 5, pp. 2686–2693, 05 1997.
[3] J.-M. Jot, “An analysis/synthesis approach to real-
time artificial reverberation,” in [Proceedings] ICASSP-
92: 1992 IEEE International Conference on Acoustics,
Speech, and Signal Processing, 1992, vol. 2, pp. 221–
224.