On the Audibility of Comb Filter DistortionsAbstractSuperpositions of delayed and undelayed versions of the same signal can occur at
different stages of the audio transmission chain. Sometimes it is a deliberate
measure to provide audio material with certain spatial or timbral qualities. Often it
is a result of multiple microphone signals, sound reflections on walls or latencies
in digital signal processing leading to comb-filter-shaped, linear distortions. The
measurement of a hearing threshold for this type of distortion with its dependance
on reflection delay, relative level and the type of audio content can be the basis for
boundaries in everyday recording practice below which undesired timbral distortions can be neglected. Therefore, a listening test was conducted to determine the
just noticeable difference for three stimulus categories (speech, a snare drum roll
and a piano phrase) and different time delays between direct and delayed signal
from 0.1 ms to 15 ms, equivalent to 0.03 - 5.15 m of sound path difference. The
results show that comb-filter distortions can still be audible if the level of the first
reflection is more than 20 dB lower than the level of the direct sound.
...
2. Listening testThe stimuli were generated using a MATLAB script. Three respresentative types of audio
signals were used: 1) a piano phrase, 2) a snare drum roll and 3) a phrase of a male speaker,
all taken from the SQUAM reference CD published by the European Broadcasting Union.
A pretest was conducted to find out the optimum duration of the stimuli to achieve the highest
sensitivity for differences between original and filtered sounds in a 3AFC situation (see
below).
An optimal stimulus duration was found out to be 3 seconds, where a minimum in the just
noticeable difference for time offsets of 0.8 and 1.2 ms and snare drum content was detected
(Fig. 4).
The main listening test started with an instruction and a training session making the subjects
familiar with the test signals and the procedure. The monophonic stimuli were presented via
headphones (Stax SR 202). The volume could be changed during the training phase to a
comfortable level and was held constant during the rest of the test. Time offsets were tested
between 0.1 and 15 ms. An 3-Down-1-Up adaptive tracking procedure was used to approximate the threshold for the just noticeable difference in a three alternative forced choice
(3AFC) test. The subject was asked to detect the stimulus which he/she thought was manipulated using a graphical user interface. The initial values for each test procedure were set
considerably higher than the threshold values determined in a pretest to allow an extra training in each test procedure. The frequency response of the comb filter was defined by the
amplitude above 0 dB at peak frequencies. For optimal stimulus placement the step size of 0.5
dB was reduced to 0.25 dB after the first reversal that led to a “down” group. The resulting
differences in loudness were compensated for using an algorithm based on Zwicker´s loudness calculation method [10]. Following a recommendation by Leek [11] the average of the
last four turnaround points was taken as the resulting threshold value.
The subjects, 36 students and professors of the TU Berlin and the UdK Berlin with experience
in audio engineering were divided into three groups of twelve. Each group listened to one of
5
24th TONMEISTERTAGUNG – VDT INTERNATIONAL CONVENTION, November, 2006
the three stimulus types (piano, snare, speech). After the test the subjects were asked to describe the cues they used to detect the distortion. The whole test took about 45 minutes per
subject.
3. ResultsThe thresholds for the three stimulus categories differ considerably.
For piano and snare drum
a minimum in the mean threshold of all subjects was found at 0.8 ms time offset. The mean
level difference between direct sound and first reflection is 13.2 dB at the threshold for piano
and 18.2 dB for snare drum, with single subjects reliably detected level differences as much as
21.5 dB (piano) and 27 dB (snare drum). The threshold curve for the speech signal shows a
trend to an easier detection of the distortion with growing time offset. For a time offset of 15
ms the just audible difference was detected at a level difference of 16 dB (average) resp. 22
dB (single subject), requiring an extension for futere tests towards higher time offsets.
6
24th TONMEISTERTAGUNG – VDT INTERNATIONAL CONVENTION, November, 2006
Fig. 5: Just noticeable difference between original and comb-filtered signal
showing the level difference between direct and delayed signal
for delay times (time offsets) between 0.1 and 15 ms.
7
24th TONMEISTERTAGUNG – VDT INTERNATIONAL CONVENTION, November, 2006
4. DiscussionThe present study proved that listeners are considerably more sensitive to frequency response
irregularities than suggested by previous investigations. For certain audio material and under
good listening conditions differences are audible when a reflection with a level difference of
18 dB (average of all listeners) respectively 27 dB (single listeners) is added. This corresponds to peaks in the comb-filter curve (Fig. 2) of 1 dB (average) and less than 0.4 dB
(single listeners). Kuhl’s finding was confirmed that listeners are particularly sensitive for
changes in timbre when dealing with noisy signals [5]. Maximum sensitivity for added reflections and broadband content is reached at time offsets between 0.1 and 1 ms, depending on
content and listener, while sensitivity with speech content rises with ascending time offset and
a minimum was not found below 15 ms.
Applied to the above mentioned occasions where comb-filter-shaped distortions can occur in
the audio transmission chain we deduce that an electrical/digital superposition of identical but
delayed content can be audible even if the delayed signal is as much as 27 dB lower than the
direct signal. In situations where the delayed signal undergoes filtering when travelling
through air oder being reflected on walls the peaks and dips in the frequency response will be
less pronounced and the established threshold can be regarded as a lower boundary. With
acoustical superpositions of sound waves a comb filter distortion will only occur for the freefield part of the sound since the diffuse-field portion looses time-coherence, thus no combfilter distortion will occur.
Even in those situations we can assume that for content with broadband frequency distribution
such as the pink-noise-like snare drum signal a distortion will be best audible for time offsets
between 0.5 and 3 ms (Fig. 5) corresponding to sound path distances between 17 cm and 1 m.
These sound path distances between microphones or at reflecting surfaces should be particularly avoided.
Further research has to be done on the perceived sound quality of added reflections. Subjects
named timbral differences to be the primary cue for the piano signal, some named the perception of a residual pitch as the primary cue for the snare drum signal, whereas for the
speech signal the primary detection cue was an impression of spaciousness even for small
time offsets.
An interesting side observation is the suggested optimal signal duration of 3 s for the recognition of frequence response irregularities in direct AB comparisons – a situation that regularly
occurs in sound engineering practice. The broad range of threshold values even for trained
listeners was unexpected, as well as the fact that the distribution for experienced sound
engineers was not much different from non-professionals among the 36 tested subjects.
Hence, audio professionals should always be aware that some of their clients might be more
sensitive to distortions than they are themselves.
