Note: Descriptions are shown in the official language in which they were submitted.
CA 02242248 1998-07-03
WO 97/32430 PCT/GB97/00432
1
This invention relates to telecommunications systems, and
is concerned in
particular with improving the quality of speech signals
transmitted over
telecommunications networks.
Signals carried over telecommunications networks are subject
to
degradation from interference, attenuation, data compression,
packet loss,
limitations in digitisation processes and other problems.
It is desirable to monitor
signals at intermediate points in their transmission paths
to identify any
imperfections and, if possible, to "repair", the signal;
that is, to restore the signal
to its original state. The "repaired" signal can then be
retransmitted. The process
can be repeated as often as necessary, according to the
length of the transmission
path and the degree of degradation, provided that at each
stage the signal has not
degraded to the point where it is no longer possible to
discern its original content.
Data signals are comparatively easy to repair as they comprise
a limited
number of characters: (e.g. binary 1 s and Os; the twelve-character
DTMF (dual
tone multiple frequency) system, or the various QAM (quadrature
amplitude
modulation) constellations. Repair of such signals can be
carried out by identifying
which of the "permitted" characters is closest to the degraded
one actually
received, and transmitting that character. For example,
in a binary system, any
signal value exceeding a threshold value may be interpreted
as a "1 ", and any
below the threshold as a "0". Check digits and other means
may be included in
the transmission to further improve the integrit
of the tra
i
i
y
nsm
ss
on.
However, in general speech signals do not have a limited
character set of
this kind, and it is thus more difficult to identify automatically
whether the signal
has been degraded at all, still less how to restore the
original signal.
In a public switched telecommunications system inter-operability
requires
that all parts of the system work compatibly. In general
this precludes complex
coding processes, at least at the interfaces between one
operator's system and
another's.
!n certain specialised applications speech signals can be transmitted as a
series of coefficients from a linear predictive coding (LPG) process, a
process
which models the excitation of a human vocal tract. These coefficients, when
applied to a vocal-tract emulating filter, can reproduce the original speech.
An
' 'E1I1 '97 12.46 u:\patents\word\25160wo.doc CA 02242248 1998-07-03 , , ,
2
example is described in US Patent 4742550 (Fette). Such a system is used, for
example, in the speech codecs (coder/decoders) used in the air interface of
mobile
telephone systemsin order to reduce the required bandwidth. However, the
transmission of speech in this form requires that specialised equipment is
present
at both transmission and receiving locations, (e.g. the mobile telephone and
radio
base station) and is thus not suitable for general use in a public switched
telecommunications network.
A number of prior-art systems are known which are arranged to identify
certain characteristics of acoustic or signal-distorting noise, and eliminate
such
characteristics. An example is disclosed in US Patent 5148488 (Chen), in which
the speech-like characteristics of the incoming signal are estimated and used
to
generate a Kalman filter. This filter is then applied to the signal, allowing
only the
speech-like properties of the received signal to pass. However, such systems
merely remove unspeechlike parts of the signal. If parts of the signal have
been
lost, or have been distorted to unspeechlike forms, they can do nothing to
restore
them.
According to a first aspect of the invention there is provided a method of
restoring a degraded speech signal received over a telecommunications system
to
an estimation of its original form, comprising the steps of:
analysing the signal according to a spectral representation model to
generate output parameters indicative of the speech content of the signal;
regenerating a speech signal derived from the output parameters so
generated; and
applying the resulting speech signal to an input of the communications
system.
According to a second aspect of this invention there is provided an
apparatus for restoring a degraded speech signal, received over a
telecommunications system to an estimation of its original form, the apparatus
comprising:
analysing means for analysing the signal using a spectral representation to
generate output parameters indicative of the speech content of the signal; and
means for generating an output signal derived from the output parameters
for regenerating the speech signal.
AME~ED SHEET
11/1 X97 12:46 u:\patents\word\2516owo.doc CA 02242248 1998-07-03 ' . . _
3
Preferably the spectral representation model is a vocal tract model, and
the regeneration of a speech signal is made using a vocal tract model.
Preferably
the regeneration model includes temporal characteristics of the regenerated
signal
which are constrained to be speech-like.
The invention, in a further aspect, also extends to a telecommunications
system having one or more interfaces with further telecommunications systems,
in
which each interface is provided with such apparatus for analysing and
restoring
signals entering and/or leaving the system.
Embodiments of the invention will now be described, by way of example
only, with reference to the accompanying drawings, in which:
Figure 1 shows a telecommunications network incorporating the invention;
Figure 2 shows a speech regeneration unit, illustrating the manner in
which an estimated "original signal" may be regenerated from a degraded input
signal;
Figure 3 illustrates a matching technique forming part of the process
employed by the speech regeneration unit of Figure 2; and
Figure 4 shows a speech regeneration unit according to the invention.
ANIE~9E0 SH~.~
CA 02242248 1998-07-03
WO 97/32430 PCT/GB97/00432
4
A description of the functional blocks in Figures 1 and 2 is given below,
and includes references to established examples of each process.
Figure 1 illustrates a generalised telecommunications system 8 comprising
a number of interconnected switches 9a, 9b, 9c, 9d, and interfacing with a
number of other systems 2a, 2b, 2c, 2d. As shown ilEustratively in Figure 2
these
may be private systems, connected to the system 8 through a private branch
exchange (PBX) 2a, international networks connected to the system 8 by way of
an International Switching Centre (ISC) 2b, another operator's public network
2c,
or another part 2d of the same operator's network. Speech signals generated at
T 0 respective sources 1 a, 1 b, 1 c, 1 d may be corrupted by the systems 2a,
2b, 2c,
2d. Speech signals entering or leaving the system 8 from or to the other
systems
2a, 2b, 2c, 2d are passed through respective speech regenerators 10a, 10b,
10c,
10d. As shown, an individual operator may choose to "ring fence" his system 8
so
that any signal entering the system 8 from another system 2a, 2b, 2c is
repaired
at the first opportunity, and any degradations to a signal are removed before
it
leaves the system. in a large network further speech regenerators (such as
regenerator 10d) may be located within the network, thereby subdividing one
operator's network into several smaller networks, 2d, 8, connected by such
speech repair units.
The system to be described only handles speech signals. If the system is
to be capable of handling data (e.g. facsimile) signals as well, separate
means (not
shown) would be necessary to identify the type of signal and apply different
restoration processes, if any, to each type. Speech/data discriminators are
well
known in the art. For example ACME (digital circuit multiplication equipment),
which uses speech compression, is provided with means for identifying the
tonal
signature of a facsimile transmission, and signals the equipment to provide a
clear
(uncompressed) transmission channel. As already indicated, data restoration
processes are commonplace in the art, and will not be described further
herein.
Figure 2 shows the general arrangement of a speech regeneration unit 1 O,
corresponding to any one of the units 10a, 10b, 1 Oc, 10d in Figure 1.
Similarly the
signal input 1 and system 2 in Figure 2 correspond to any 'one of the inputs 1
a, 1 b,
1 c, 1 d and their respective systems 2a, 2b, 2c or 2d.
CA 02242248 1998-07-03
WO 97/32430 PCT/GB97i00432
The signal input 1 provides the original speech material received by the
first telecommunications system 2. This material may be transmitted over part
of
the system 2 in a digital form, but the signal to be analysed is an analogue
signal.
This analogue signal is a degraded form of the original analogue speech
signal; the
5 degradations being due to the factors referred to
previously, including the
digitisation process itself. The analogue speech signal is output from the
system 2
to the speech regenerator 10. in the regenerator 10 the distorted speech
signal is
first passed to a speech recogniser 3 which classifies the distorted speech
sound,
to facilitate selection of an "original sound" file from a memory store of
such files
forming part of the recogniser 3.
In this specification the term "speech recognition" is used to mean the
recognition of speech events from a speech signal waveform. In the area of
speech
technology, the use of machines to recognise speech has been the goal of
engineers
and scientists for many years. A variety of practical speech recognisers have
appeared in the literature including description of; HMM (Hidden Markov
Models) Cox
1990: [Wheddon C and Linggard R: "Speech communication", Speech and Language
Processing, Chapman and Hall ( 1990)] fixed dimension classifiers (such as
nearest
neighbour, Gaussian mixtures, and multi-layer perception) [Woodland & Miliar
1990:
ibid], and neural arrays [Tattersall, Linford & Linggard 1990: ibid].
Most recognition systems consist of a feature extractor and a pattern
matching process (classification) and can be either speaker-dependent or
speaker-
independent. Speaker-dependent recognisers are trained by the user with each
of
the words required for the particular application. Speaker-independent
recognition
systems have a prescribed vocabulary which cannot be changed [Wheddon C &
Linggard R: "Speech communication", Speech and Language Processing, Chapman &
Hall (1990)]. fn both systems features are extracted from the acoustic signal
which
are passed to a classifier which determines which of the words in its
vocabulary was
spoken. Features are extracted using transform or digital filtering techniques
to
reduce the amount of data passed to the classifier. The resulting patterns are
then
warped in time to optimally align with the reference patterns jSakoe H and
Chibass:
"Dynamic programming algorithm optimisation for spoken word recognition", IEEE
Trans Acoust Speech and Signal Proc, 26 (1978)). Statistical models such as
hidden
Markov models [Cox S J: "Hidden Markov models for automatic speech
recognition:
theory and application", BT Telecom Technol J, 6, No. 2 (1988)] are also
widely
CA 02242248 1998-07-03
WO 97/32430 PCT/GB97/00432
6
used. Here a sequence of features is compared with a set of probabilistically
defined
word models. Feature extraction and pattern matching techniques may also be
extended to cope with connected words EBridle J S, Brown M D and Chamberlain R
M: "An algorithm for connected word recognition", Automatic Speech Analysis
and
Recognition, Reidal Publishing Company (1984)] which is a far more complex
task as
the number of words is unknown and the boundaries between words cannot be
easily determined in real time. This results in increased computation time
EAtai B S
and Rabiner L R: "Speech research directions", AT&T Technical Journal 65,
Issue 5
(1986)] and a corresponding increase in hardware complexity.
Hidden Markov Models suitable for the present purpose are described in
Baun L E, "An lnegualiry and Associated Maximisation Technique in Statistical
Estimation for Probabilistic Functions of Markov Processes" Inequalities II1,
1-8,
1972, or Cox S J, "Hidden Markov Models For Automatic Speech Recognition;
Theory and Application", in "Speech and Language Processing" edited by
Wheddon C and Linggard R, Chapman and Hall, ISBN O 412 37800 0, 1990. The
HMM represents known words as a set of feature vectors, and, for a given
incoming word, calculates the a posteriori probability that its model will
produce
the observed set of feature vectors. A generic "original sound" file can then
be
selected from memory for the recognised word.
The "original sound" file so identified is then used to control a speech
generator 7 to generate an audio signal corresponding to the sound to be
produced. Thus the speech recogniser identifies which speech element was the
most likely to have been present in the original signal, and the speech
generator
then produces an undistorted version of that speech element, from a store of
such
speech elements. The output thus consists only of speech-like elements.
Provided
that the signal received from the telecommunications system is not so
corrupted
that the speech recogniser 3 fails to identify the correct speech element, the
output from the speech generator 7 should be purely the speech content of the
original signal.
The macro properties of the synthesised speech generated by the
generator 7 are now adapted in an adaptor 4 to those of the actual speech
event.
The adaptor 4 reproduces the characteristics of the original talker,
specifically
fundamental frequency (which reflects the dimensions of the individual's vocal
CA 02242248 1998-07-03
WO 97!32430 PCT/GB97i00432
7
tract), glottal excitation characteristics, which determine
the tonal quality of the
voice, and temporal warping, to fit the general template
to the speed of delivery of
the individual speech elements. This is to allow the general
"original sound" file to
be matched to the actual speech utterances, making the technique
practically
robust, and talker-independent. These characteristics are
described in
"Mechanisms of Speech recognition", W,A, Ainsworth, Pergamon
Press, 1976.
The pitch /fundamental frequency) of the signal may be matched
to that
of the stored "original sound", by matching the fundamental
0 frequency of each
output element, or some other identifiable frequency, to
that of the original voice
signal so as to match the inflections of the original speaker's
voice.
Glottal excitation characteristics can be produced algorithmically
from
analysis of the characteristics of the original signal,
as described with reference to
Figure 4.3 (page 36) of the Ainsworth reference cited above.
The mathematical technique used for time warping is described
for
example in Holmes J N, "Speech Synthesis and Recognition",
Van Nostrand
Reinhold (UK) Co. Ltd., fSBN 0 278 00013 4, 1988, and Bridle
0 J S, Brown M D,
Chamberlain R M, "Continuous Connected Word Recognition
Using Vllhoie Word
Templates", Radio and Electronics Engineer 53, Pages 167-177,
1983. The time
5 alignment path between the two words /uttered and recognised
"original"), see
Figure 3, describes the time warping required to fit the
stored "original sound" to
that of the detected word. Figure 3 shows, on the vertical
axis, the elements of
the recognised word "pattern", and on the horizontal axis
the corresponding
elements of the uttered word. It will be seen that the speaker's
utterance differs
from the word retrieved from the store in the length of
certain elements and so, in
order to match the original utterance certain elements,
specifically the "p" and "r",
are lengthened and others, specifically the "t", are shortened.
The regenerated signal is then. output to the telecommunications
system 8.
Although the speech recogniser 3, speech generator 7 and
30 adaptor 4 have
been described as separate hardware, in practice they could
be realised by a single
suitably programmed digital processor.
The above system requires a large memory store of recognisable
speech
words or word elements, and will only reproduce a speech
element if it recognises
it from its stored samples. Thus any sound produced at the
output of the
1 1 /1 ', X97 1 2:46 u:\patents\word\251 60wo.doc
CA 02242248 1998-07-03 ' '
., ,
", ,
i
8
telecommunications system 2 which is not matched with one stored in the
memory will be rejected as not being speech, and not retransmitted. In this
way,
only events in the signal content recognised as being speech will be
retransmitted,
and non-spEech events will be removed.
In an embodiment of the invention, shown in Figure 4, the speech
regeneration unit is made up of a vocal tract analysis unit 1 1 , the output
of which
is fed to a vocal tract simulator 12 to generate a speech-like signal. This
system
has the advantage that non-speech-like parameters are removed from otherwise
speech-like events, instead of each event being accepted or rejected in its
entirety.
The vocal tract analysis system stores the characteristics of a generalised
natural system (the human vocal tract) rather than a "library" of sounds
producable
by such a system. The preferred embodiment of Figure 4 therefore has the
advantage over the embodiment of Figure 2 that it can . reproduce any sound
producable by a human vocal tract. This has the advantages that there is no
need
for a large memory store of possible sounds, nor the consequent processing
time
involved in searching it. Moreover, the system is not limited to those sounds
which
have been stored.
It is appropriate here to briefly discuss the characteristics of vocal tract
analysis systems. The vocal tract is a non-uniform acoustic tube which extends
from the glottis to the lips and varies in shape as a function of time [Fant G
C M,
"Acoustic Theory of Speech Production", Mouton and Co., 's-Gravehage, the
Netherlands, 1960]. The major anatomical components causing the time varying
change are the lips, jaws, tongue and velum. For ease of computation it is
desirable that models for this system are both linear and time-invariant.
Unfortunately, the human speech mechanism does not precisely satisfy either of
these properties. Speech is a continually time varying-process. In addition,
the
glottis is not uncoupled from the vocal tract, which results in non-linear
characteristics [Flanagan J L "Source-System Interactions in the Vocal Tract",
Ann. New York Acad. Sci 155, 9-15, 1968]. However, by making reasonable
assumptions, it is possible to develop linear time invariant models over short
intervals of time for describing speech events [Market J D, Gray A H, "Linear
Prediction of Speech", Springer-Verlag Berlin Heidelberg New York, 1976].
Linear
predictive codecs divide speech events into short time periods, or frames, and
use
AMENDED SHEEN
CA 02242248 1998-07-03
WO 97/32430 PCT/GB97/00432
9
past speech frames to generate a unique set of predictor parameters to
represent
the speech in a current frame [Atal B S, Hanauer S L "Speech Analysis and
' Synthesis by Linear Prediction of the Speech Wave" J. Acoust. Soc. Amer.,
vol.
50, pp. 637-655,1 971 1. Linear predictive analysis has become a widely used
method for estimating such speech parameters as pitch, formants and spectra.
Auditory models (timeifrequency/amplitude spectrograms) rely on audible
features
of the sound being monitored, and take no account of how they are produced,
whereas a vocal tract model is capable of identifying whether the signal is
speech-
like, i.e. whether a real vocal tract could have produced it. Thus inaudible
differences, not recognised by auditory models, will nevertheless be
recognised by
a vocal tract model.
A vocal tract model suitable for use in the analysis is the Linear Predictive
Coding model as described in Digital Processing of Speech Signals: Rabiner
L.R.;
Schafer R.W; (Prentice-Hall 1978) page 396.
Enhancements of the vocal tract model may include the inclusion of
permissible temporal characteristics, such as long-term pitch prediction,
which
allow the regeneration of speech components which are missing from a given
speech structure, or so badly distorted that they fail to be recognised by the
analysis process. The inclusion of such temporal characteristics would smooth
out
implausibly abrupt onsets, interruptions or ends of speech components, which
may
be caused, for example, by the brief loss or corruption of a signal.
The parameters generated by the vocal tract mode! 1 1 identify the speech
like characteristics of the original signal. Any characteristics which are not
speech
like are unable to be modelled by the vocal tract model, and will therefore
not be
parameterised.
The parameters generated by the vocal tract model are used to control a
speech production model 1 2. The parameters modify an excitation signal
generated
by the synthesiser, in accordance with the vocal tract parameters generated by
the
analyser 1 1, to generate a speech-like signal including the speech-like
characteristics of the signal received from the system 2, but not the
distortions.
Suitable vocal tract models for use in the synthesis include the Linear
Predictive Coding model described above, or a more sophisticated model such as
the cascade/paralfel formant synthesiser, described in the Journal of the
Acoustic
CA 02242248 1998-07-03
WO 97!32430 PCT/GB97/00432
Society of America (Vol 67, No3, March 19801: D.H. K(att; "Software for a
CascadelParaliel Formant Synthesiser".
a
Other suitable systems are disclosed in "Phase Coherence in Speech
Reconstruction for Enhancement and Coding Applications": Quatieri et al: _
5 International Conference on Acoustics, Speech, and Signal Processing, Vol 1
23-
26 May 1989, Glasgow (Scotland!: pages 207-210; and Kamata et al
"Reconstruction of Human Voice using Parallel Structure Transfer Function and
its
Estimation Error': IEEE Pacific Rim Conference on Communications, Computers
and
Signal Processing; 17-19 May 1995 Victoria, British Columbia, Canada.
10 It should be understood that the term "speech", as used in this
specification, is used to mean any utterance capable of production by the
human
voice, including singing, but does not necessarily imply that the utterance
has any
intelligible content.
