Fairbanks, Grant. Experimental Phonetics – T04

Method for Time or Frequency
Compression-Expansion of Speech *1 ***2

Grant Fairbanks, W. L. Everitt, and R. P. Jaeger **3
University of Illinois, Urbana, Illinois

The purposes of this paper are to outline a method
for compression and expansion of speech, to describe the
device employed in the method, and to demonstrate by
means of recordings the results of the method at this
experimental stage.

Until comparatively recently we had not been aware
of the fact that several approaches to the problem similar
to ours had previously been made by other experimenters.
We have now learned that our method, although
developed independently, resembles in certain
features of theory and details the earlier work of French
and Zinn, 14 Gabrilovitch, 15 Haase, 16 Gabor, 27 Vilbig, 38 and,
perhaps, others.

Fundamentally, the process depends upon the fact
that the duration of the average speech element or
phoneme of live connected speech, such as ah or s or r,
exceeds the minimum duration necessary for perception
by a listener, or exceeds the minimum time necessary for
sampling the essential phonemic qualities of the speech
element in question. This minimum duration has been
the object of a psychophysical study by Peterson 49 and of
theoretical calculation by Gemelli and Pastori. 510 The
excess duration may be referred to as temporal redundancy,
which term we suggest as a useful specification
at the experimental level when spoken language is in
question.

The dimensions of the problem are clearly not only
those of engineering, but also those of psychophysics. In
this paper we confine ourselves to the method. A psychophysical
program is in progress and its results will
be reported separately.

For purposes of explanation assume two different
phonemes, A and B, which are of equal duration and
joined without interruption as shown (Fig. 1). Assume
that Aʼ and Bʼ are valid samples of A and B, and that
each is of adequate duration for perception. Assume
that samples Aʼ and Bʼ are extracted from A and B and
abutted in time as shown without discontinuity, and that
A — Aʼ and B — Bʼ are discarded. If, now, Aʼ, Bʼ is
reproduced, the time will be shorter than the original A,
B, but the phonemes should be perceptible.

In the case of expansion, assume that phonemes A
and B are caused to be repeated, as in the middle portion.
If A, A, B, B is reproduced, the time will be longer
and the auditory effect, given the above assumptions,
should be that of prolongation of A and of B.

Finally, assume that A and B are first compressed
to Aʼ and Bʼ, and then expanded to Aʼ, Aʼ, Bʼ, Bʼ as
shown at the bottom. Here the original time for A and
B has been restored. A and B have been reconstructed
from Aʼ and Bʼ.

Fig. 2 shows a photograph of the essential part of
an experimental model of a device for compression or
expansion along the lines of such a theory. Basically,
the device is a continuous loop magnetic tape recorder,
mounted at the bottom of the rack containing the other
components. The tape loop, approximately 12 feet long,
rises along the right edge of the rack to a pulley under
slight spring tension at the top. Its pathway is shown
by arrows. Entering the device, the tape is directed by
means of rollers over a Magnecord erase head, and then
over a fixed Magnecord record head where the input is
temporarily recorded. Passing over another roller, the
tape then descends to a revolving playback head assembly
enclosed in a mu-metal box, where the signal recorded
on the loop is scanned. Next the tape passes to
the drive capstan, around a roller, and, finally, over a
Brush permanent magnet erase head.

The revolving head assembly consists of a brass drum
with four Brush playback heads equally spaced around
its periphery. The output of the heads is taken off by
means of a slipring-brush unit The circumferences of
both drum and capstan are 7.64″. Drum and capstan
are mounted on shafts supported in sleeve bearings
at the back of the panel. Massive flywheels are also
mounted on the shafts. The two units are driven by
twin 1/15 hp DC Bodine motors with independent speed
38controls by means of GR Variacs. Speeds are measured
with a GR Strobotac.

The remaining components are conventional. An
independent Magnecorder PT6-A is used for storage
and playback. This has been modified for continuously
variable speed reduction and furnished about a 15 to 1
range of tape velocities.

In Fig. 3 operation of the revolving head assembly
is shown at the left. The four playback heads are identified
by letters. The tape passes over the drum and is in
contact with ¼ of its circumference, or a distance equal
to the peripheral distance between any two adjacent
playback heads. The tape is retained by flanges around
the drum periphery. Tape direction is constantly
counter-clockwise. In the compression application the
direction of drum rotation is also counter-clockwise.
Under load the top tape velocity is approximately 190
in/sec. The top peripheral drum velocity is about 225
in/sec.

For purposes of explanation the tape is divided into
hypothetical numbered segments, each equal to the distance
between heads. The relative positions of tape and
heads are shown at representative times. The diagram
shows 50 per cent time compression as an example.

At time II head A is at 6 o'clock and head B is at
9 o'clock, while segment 2 lies between them in contact
with the drum. Head A is about to leave the drum,
while head B is about to begin reproducing segment 3.
Accordingly, although there is no discontinuity, segment
2 is not reproduced by any head. The remaining diagrams
show how the process continues, the odd-numbered
segments being reproduced at reduced frequency
and the even-numbered segments being discarded. It is
evident that various durations of either reproduced or
discarded segments can be realized by varying the absolute
and relative velocities of tape and head, and that
a range of sampling frequencies and compression ratios
can thus be produced.

With respect to duration the odd-numbered segments
are termed sampling intervals; the even-numbered segments
discard intervals. The reciprocal of their summed
durations is the sampling frequency. One hundred times
the discard interval divided by the sum of the two intervals
will be termed the compression percentage. Since
sampling is periodic the ratio applies also to the total
message time, and describes the percentage by which
that total time has been reduced.

Assuming that the process results in intelligible speech,
it becomes evident that the processed message may be
transmitted over a system with smaller bandwidth than
originally necessary. The capacity of a conventional
transmission link for handling simultaneous messages
will be a function of the amount of compression, or frequency
division.

Fig. 4 is a similar diagram for expansion. Here the
drum bearing the playback head revolves in a direction
opposite to that of the tape. The illustrative example
shows the condition when these velocities are equal.
The effective velocity is equal to their sum.

At t₀, shown at I, segment 1 is in contact with the
drum between heads A and D. During the next interval
head D, as it moves from 6 o'clock to 9 o'clock, will
reproduce both segments 1 and 2 and then leave the
tape. At that time it will be replaced by head C, which
has moved to the 6 o'clock position to intercept the tape
at the beginning of segment 2, and which will reproduce
segments 2 and 3 during its sweep. The result, shown
at the right, is that between times I and II, while segments
1 and 2 are passing the 6 o'clock point, segments
1, 2, 2, 3 are reproduced. The rest of the figure shows
how this process continues.

Since the effective tape velocity has been increased
by the opposite movement of head and tape, frequency
multiplication has been incurred. The original frequencies
are restored by reproducing the processed message
in an appropriately longer time. One hundred times the
amount of time thus added divided by the original time
is the expansion percentage. In the diagram this equals
100 per cent.

Note that small percentages did not affect the perceived
speed of talking very much, and that the details
of speech became more readily heard as expansion increased.
Toward the end you may have heard an echolike
sound. This occurs when the interval repeated
exceeds the duration of one phoneme. This is a size
limitation in our experimental model and not a limitation
of the method.

The process is illustrated in the next recordings. First
you will hear the original message. Then you will hear
the transmitted message with frequency division. Finally
you will hear the message as received after reconstruction
by means of expansion and corresponding frequency
multiplication. Eighty per cent of the message was discarded
before transmission and the final message as
you hear it was reconstructed from the 20 per cent fragment
that remained. To help you appreciate the last
point we will also “play” at the end a recording in
which the original frequencies are restored by accelerated
playback without time expansion.

We present this next recording with some hesitation
and we hope you will not be disappointed. It was made
on an experimental model of the device. Its main purpose
is to validate the theory and demonstrate potential
feasibility. (You will “hear” considerable noise and distortion.
Some of this can be eliminated fairly readily,
but part of it is inherent in the method and will need
to be counteracted.)

Apart from its theoretical interest, the method appears
to have several practical applications. For one
thing, the smaller compression and expansion ratios
should be useful in the programming of rebroadcast
speeches in radio, since they furnish “tailormade” time
without the audience's knowledge. A saving of 10 minutes
per hour is completely realistic. Conversely, and we
advance this suggestion with diffidence, thinking of
commercials, more intelligence can be communicated to
an audience in a given amount of time.

Straightforward compression by larger amounts should
be useful wherever high-speed communication is crucial,
as in certain military situations. Expansion should facilitate
branches of study such as experimental phonetics
and linguistics where auditory analysis is important.

Finally, of course, the method gives promise as an approach
to the long-standing problem of bandwidth reduction.