 |
Description  |
|
|
BACKGROUND OF THE INVENTION
The technical field of this invention is speech transmission and, in
particular, methods and devices for pre-processing audio signals prior to
broadcast or other transmission.
The problem of speech degradation by natural or man-made disturbances is
one which commonly occurs in AM radio broadcasting and ground-to-air
communications. Often in these applications, a peak-power limitation is
imposed by the transmitter or a dynamic range constraint results either
from the sensitivity characteristics of the receiver or from the ambient
noise level. Under these constraints, the audio signals are preprocessed
to increase intelligibility. Techniques such as dynamic range compression,
pre-emphasis and clipping have been applied with limited success to reduce
the peak factor of a waveform in order to increase loudness while
attempting to preserve important features of the spectral envelope. For a
further description of such techniques, see Modulation-Process Techniques
for Sound Broadcasting, Tech. 3243-E, Technical Center of the European
Broadcasting Union, Bruxelles, Belgium, July 1985, herein incorporated by
reference.
There exists a need for better preprocessing techniques for speech
transmission, particularly where the spectral magnitude is specified and
the goal is to achieve a flattened time-domain envelope which satisfies
peak power limitations. In particular, new techniques for accomplishing
automatic gain control, (multiband) dynamic range compression,
pre-emphasis and phase dispersion would satisfy a long-felt need in the
field.
The above-referenced parent application U.S. Ser. No. 712,866 discloses
that speech analysis and synthesis as well as coding and time-scale
modification can be accomplished simply and effectively by employing a
time-frequency representation of the speech waveform which is independent
of the speech state. Specifically, a sinusoidal model for the speech
waveform is used to develop a new analysis-synthesis technique.
The basic method of U.S. Ser. No. 712,866 includes the steps of: (a)
selecting frames (i.e. windows of about 20-40 milliseconds) of samples
from the waveform; (b) analyzing each frame of samples to extract a set of
frequency components; (c) tracking the components from one frame to the
next; and (d) interpolating the values of the components from one frame to
the next to obtain a parametric representation of the waveform. A
synthetic waveform can then be constructed by generating a series of sine
waves corresponding to the parametric representation. The disclosures of
U.S. Ser. No. 712,866 are incorporated herein by reference.
In one illustrated embodiment described in detail in U.S. Ser. No. 712,866,
the basic method summarized above is employed to choose amplitudes,
frequencies, and phases corresponding to the largest peaks in a
periodogram of the measured signal, independently of the speech state. In
order to reconstruct the speech waveform, the amplitudes, frequencies, and
phases of the sine waves estimated on one frame are matched and allowed to
continuously evolve into the corresponding parameter set on the successive
frame. Because the number of estimated peaks are not constant and slowly
varying, the matching process is not straightforward. Rapidly varying
regions of speech such as unvoiced/voiced transitions can result in large
changes in both the location and number of peaks. To account for such
rapid movements in spectral energy, the concept of "birth" and "death" of
sinusoidal components is employed in a nearest-neighbor matching method
based on the frequencies estimated on each frame. If a new peak appears, a
"birth" is said to occur and a new track is initiated. If an old peak is
not matched, a "death" is said to occur and the corresponding track is
allowed to decay to zero. Once the parameters on successive frames have
been matched, phase continuity of each sinusoidal component is ensured by
unwrapping the phase. In one preferred embodiment the phase is unwrapped
using a cubic phase interpolation function having parameter values that
are chosen to satisfy the measured phase and frequency constraints at the
frame boundaries while maintaining maximal smoothness over the frame
duration. Finally, the corresponding sinusoidal amplitudes are simply
interpolated in a linear manner across each frame.
SUMMARY OF THE INVENTION
A sinusoidal speech representation system is applied to the problem of
speech dispersion by pre-processing the waveform prior to transmission to
reduce the peak-to-RMS ratio of the waveform. The sinusoidal system first
estimates and then removes the natural phase dispersion in the frequency
components of the speech signal. Artificial dispersion based on pulse
compression techniques is then introduced with little change in speech
quality. The new phase dispersion allocation serves to preprocess the
waveform prior to dynamic range compression and clipping, allowing
considerably deeper thresholding than can be tolerated on the original
waveform.
Whereas conventional systems accomplish phase dispersion using all-pass
dispersion networks, it is shown that, using the sinsoidal system, the
phases of the individual sine waves can be manipulated to achieve
improvements in the peak-to-RMS ratio. For example, dispersion of the
speech waveform can be performed by first removing the vocal tract system
phase derived from the measured sine-wave amplitudes and phases, and then
modifying the resulting phase of the sine waves which make up the speech
vocal cord excitation.
The present invention also allows for (multiband) dynamic range
compression, pre-emphasis and adaptive processing. A method of dynamic
range control is described, which is based on scaling the sine-wave
amplitudes in frequency (as a function of time) with appropriate attack
and release-time dynamics applied to the frame energies. Since a uniform
scaling factor can be applied across frequency, the short-time spectral
shape is maintained. The phase dispersion solution can also be applied to
determine parameters which drive dynamic range compression and, hence, the
phase dispersion and dynamic range procedures can be closely coupled to
each other. In addition, the sinusoidal system allows dynamic range
control to be applied conveniently to separate frequency bands, utilizing
different low- and high-frequency characteristics. Pre-emphasis, or any
desired frequency shaping, can be performed simply by shaping the
sine-wave amplitudes versus frequency prior to computing the phase
dispersion. The phase dispersion techniques can take into account and
yield optimal solutions for any given pre-emphasis approach.
The sinusoidal analysis/synthesis system is also particularly suitable for
adaptive processing, since linear and non-linear adaptive control
parameters can be derived from the sinusoidal parameters which are related
to various features of speech. For example, one measure can be derived
based on changes in the sinusoidal amplitudes and frequencies across an
analysis frame duration and can be used in selectively accentuating
frequency components and expanding the time scale.
The invention will next be described in connection with certain illustrated
embodiments. However, it should be clear that various modifications,
additions and subtractions can be made by those skilled in the art without
departing from the spirit and scope of the invention.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a flow diagram of a method for introducing an artificial phase
dispersion according to the present invention.
FIG. 2 is a general block diagram of an audio pre-processing system
according to the present invention.
FIG. 3 is a more detailed illustration of the system of FIG. 2.
FIG. 4 is a more detailed illustration of the phase dispersion computer of
FIG. 3.
DETAILED DESCRIPTION
In FIG. 1, a schematic approach according to the present invention is shown
whereby the natural dispersion of speech is replaced by a desired
dispersion which yields a pre-processed waveform suitable for dynamic
range compression and clipping prior to broadcast or other transmission to
improve range and/or intelligibility. The object of the present invention
is to obtain a flattened, time-domain envelope which can satisfy peak
power limitations and to obtain a speech waveform with a low peak-to-RMS
ratio.
In FIG. 2, a block diagram of the audio preprocessing system 10 of the
present invention is shown consisting of a spectral analyzer 12,
pre-emphasizer 14, dispersion computer 16, envelope estimator 18, dynamic
range compressor 20 and waveform clipper 22. The spectral analyzer 12
computes the spectral magnitude and phase of a speech frame. The magnitude
of this frame can then be pre-emphasized by pre-emphasizer 14, as desired.
The system (i.e., vocal tract) contributions are then used by the
dispersion computer 16 to derive an optimal phase dispersion allocation.
This allocation can then be used by the envelope estimator 18 to predict
an time-domain envelope shape, which is used by the dynamic range
compressor 20 to derive a gain which can be applied to the sine wave
amplitudes to yield a compressed waveform. This waveform can be clipped by
clipper 22 to obtain the desired waveform for broadcast by transmitter 24
or other transmission.
In FIG. 3, the system 10 for pre-processing speech is shown in more detail
having a Fast Fourier Transformer (FFT) spectral analyzer 12, system
magnitude and phase estimator 34, an excitation magnitude estimator 36 and
an excitation phase estimator 38. Each of these components can be similar
in design and function to the same identified elements shown and described
in U.S. Ser. No. 712,866. Essentially, these components serve to extract
representative sine waves defined to consist of system contributions
(i.e., from the vocal tract) and excitation contributions (i.e., from the
vocal chords). Similarly, a peak detector 40 and frequency matcher 42,
along the same lines as those described in U.S. Ser. No. 712,766 are
employed to track and match the individual frequency components from one
frame to the next. A pre-emphasizer 14, also known in the art, can be
interposed between the spectral analyzer 12 and the system estimator 34.
In a simple embodiment, the speech waveform can be digitized at a 10 kHz
sampling rate, low-passed filtered at 5 kHz, and analyzed at 10 msec frame
intervals with a 25 msec Hamming window. Speech representations, according
to the invention, can also be obtained by employing an analysis window of
variable duration. For some applications, it is preferable to have the
width of the analysis window be pitch adaptive, being set, for example, at
2.5 times the average pitch period with a minimum width of 20 msec.
To achieve continuity at the frame boundaries, the magnitude and phase
values must be interpolated from frame to frame. The system magnitude and
phase values, as well as the excitation magnitude values, can be
interpolated by linear interpolator 44, while the excitation phase values
are preferably interpolated by cubic interpolator 46. Again, this
technique is described in more detail in parent case, U.S. Ser. No.
712,866, herein incorporated by reference.
The illustrated system employs a pitch extractor 32. Pitch measurements can
be obtained in a variety of ways. For example, the Fourier transform of
the logarithm of the high-resolution magnitude can first be computed to
obtain the "cepstrum". A peak is then selected from the cepstrum within
the expected pitch period range. The resulting pitch determination is
employed by the phase dispersion computer 16 (as described below) and can
also be used by the system estimator 34 in deriving the system magnitudes.
In the system estimator 34, a refined estimate of the spectral envelope can
be obtained by linearly interpolating across a subset of peaks in the
spectrum (obtained from peak detector 40) based on pitch determinations
(from pitch extractor 32). The system estimator 34 then yields an estimate
of the vocal tract spectral envelope. For further details, again, see U.S.
Ser. No. 712,866.
In the present invention, the excitation phase estimator 38 is employed to
generate an excitation phase estimate. In one embodiment, using a Hilbert
Transform with the system amplitude, an initial (minimum) phase estimate
of the system phase is obtained. The minimum phase estimate is then
subtracted from the measured phase. If the minimum phase estimate were
correct, the result would be the linear excitation phase. In general,
however, there will be a phase residual randomly varying about the linear
excitation phase. A best linear phase estimate using least squares
techniques can then be computed. For a further discussion of excitation
phase estimation, see a paper by the present inventors "Phase Modeling And
Its Application To Sinusoidal Transform Coding" Proceedings of ICASSP
1986.
In estimating the excitation function, small errors in the linear estimate
can be corrected using the system phase. The system phase estimate can be
obtained by subtracting the linear phase from the measured phase and then
used along with the system magnitude to generate a system impulse response
estimate. This response can be cross-correlated with a response from the
previous frame. The measured delay between the responses can be used to
correct that linear excitation phase estimate. Other alignment procedures
will be apparent to those skilled in the art.
In the present invention, an artificial system phase is computed by phase
dispersion computer 16 from the system magnitude and the pitch. The
operation of phase dispersion computer 16 is shown in more detail in FIG.
4, where the raw pitch estimate from the cepstral pitch extractor 32 is
smoothed (i.e. by averaging with a first order recursive filter 50) and a
phase estimate is obtained by phase computer 52 from the system magnitude
by the following equation:
##EQU1##
where,
##EQU2##
where .theta.(.omega.) is the artificial system phase estimate and k is
the scale factor and M(.omega.) is the system magnitude estimate. This
computation can be implemented, for example, by using samples from the FFT
analyzer 12 and performing numerical integration.
The scale factor k is obtained by the scale factor computer 54 by solving
the following equation
k=2.pi.(pitch period)/g(.pi.) (2)
where g (.pi.) is the value of EQ. (1B) at .pi.. Multiplier 56 multiplies
the phase computation by the scale factor to yield the system phase
estimate .theta.(.omega.) for phase dispersion, which can then be further
smoothed along the frequency tracks of each sine wave (i.e., again using a
1st order recursive filter 58 along such frequency tracks). The system
phase is then available for interpolation.
With reference again to FIG. 2, the system phase can also be used by
envelope estimator 18 to estimate the time domain envelope shape. For
example, the envelope can be computed by using a Hilbert transform to
obtain an analytic signal representation of the artificial vocal tract
response with the new phase dispersion. The magnitude of this signal is
the desired envelope. The average envelope measure is then used by dynamic
range compressor 20 to determine an appropriate gain. The envelope can
also be obtained from the pitch period and the energy in the system
response by exploiting the relationship of the signal and its Fourier
transform. A desired output envelope is computed from the measured system
envelope according to a dynamic range compression curve and appropriate
attack and release times. The gain is then selected to meet the desired
output envelope. The gain is applied to the system magnitudes prior to
interpolation.
Alternatively, the dynamic range compressor 20 can determine a gain from
the detected peaks by computing an energy measure from the sum of the
squares of the peaks. Again, a desired output energy is computed from the
measured sinewave energy according to a dynamic range compression curve
and appropriate attack and release times. The gain is then selected to
meet the desired output energy. The gain is applied to the sinewave
magnitudes prior to interpolation.
After interpolation, sinewave generator 60 generates a modified speech
waveform from the sinusoidal components. These components are then summed
and clipped by clipper 22. The spectral information in the resulting
dispersed waveform is embedded primarily within the zero crossings of the
modified waveform, rather than the waveform shape. Consequently, this
technique can serve as a pre-processor for waveform clipping, allowing
considerably deeper thresholding (e.g., 40% of the waveform's maximum
value) than can be tolerated on the original waveform.
* * * * *
|
|
 |
|
 |
Description |
|
