Note: Descriptions are shown in the official language in which they were submitted.
- ~08~3~
METHOD FOR THE QUANTIFICATION OF THE ENERGY OF THE
SPEECH SIGNAL IN A VOCODER WITH VERY LOW BIT RATE
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to a method for the
quantification of the energy of the speech signal in a
vocoder with a very low bit rate.
It can be applied notably to the making of the
linear predlction vocoders used for the transmission of
speech by radio, similar to those described for example
in the Revue Technique THOMSON-CSF (THOMSON-CSF
Technical Journal), volume 14, No. 3, September 1982,
pp. 715 to 731, in which the~ speech signal is
identified at the output of a digital filter, the input
of which recelves either a periodic waveform
corresponding to the waveforms of its voiced sounds
such as the vowels or a random waveform corresponding
to the waveforms of its unvoiced sounds such as most of
its consonants.
2. Description of the Prior Art
. .
It is known that the auditory quality of linear
prediction vocoders depends greatly on the precision
with which their predictive filter is quantified, but
also on the quality of the restitution of the power
profile of the excita~ion. This is especially true for
certain transitory sounds such as many consonants: for
.
.
~ '
example, poor quality restitution does not allow a "d"
to be distinguished from a "t" or from a "k".
As a rule, the speech signal is segmented into
frames of constant duration, and a single value of
power (or energy) is given ~ox each frame.
In vocoders with very low bit rate, one way to
lower the bit rate is to increase the duration of the
frame, for example from 22.5 ms to 30 ms as well as to
group together and quantify the parameters relating to
several frames once alone. This enables the dlfferent
parameters of synthesis to be renewed less frequently.
Unfortunately, the intelligibility of the restituted
speech is diminished, for the transmitting of only one
value of ener~y per frame no longer enables the
appropriate restitution of certain transitory sounds.
A first known way to overcome these difficulties
consists in grouping the frames together in packets
while considering k values o~ energy per packet, each
of which can be represented by the coordinates of a
point referenced in a k-dimensional space. A
statistical analysis makes it possible to determine the
main axas of the cloud of the poin~s observed. The
quantification takes place on the coordinates of the
points borne by the main axes t each point being
quantified on a number o~ bits depending on the eigen
value or characteristic value associated with each axis
considered. However, he drawback of operating in this
~3~
way is that it is necessary to plan a p.rocedure of
correction at the synthesis filter so that the values
of the energies compute~ are not negative. Furthermore,
in this processing operation, no special attention is
paid to the fidelity of restitution of the transitory
sounds.
According to a second method, also known, which
partly follows the procedure of the first method by the
grouping of frames in packets and which also takes k
values of energy per packet into consideration, the k
values of energy are no longer encoded in a scalar way
but vectorially by means of a dictionary containing M =
2Q multiplets of k v~lues each in considering the k
values to be quantified on Q bits.
In this case, the difficulties of setting up the
system appear from the fact that it is necessary,
firstly, to create and store a dictionary and,
secondly, to carry out a quantification. Since the
dictionary is generally poorly structured and since it
is necessary to count at least two bits per value of
energy, the encoding o~ the number Q occupies no less
than 22 combinations which represents very major
computing loads for the signal processors of the
vocoders.
SUMMARY OF THE INVENTIO~
:
It is the aim of :the invention to overcome the
above-mentioned drawbacks. To this effect, an object of
'.
:
,
- 2~833~
the invention is a method ~or the quan-tification of the
energy of the speech signal in a vocoder with very low
bit rate, said method consisting in dividing (1) the
speech signal into packets of a determined number of
~frames of a constant duration by the sampling of a
determined number n o~ energy values in each frame,
quantifying ~2, 3, 4) the first energy value measured
in each fixst frame of a packet according to a
determinéd number QO of bits and the variations of the
k - 1 remaining energies in relation to the first value
of the energy sampled on a determined number Q1 of bits
smaller t,han Q0, the variations~ of the k - 1 energies
being selected from a table of "slopes" enabling each
energy sample k ~o be assigned the energy "slope'i that
separates~i~ from the energy of the "k - 1th" or "k - 1
order" previous sample.
The main advantage of the method according to the
invention is that it can be used to obtain high quality
energy in each frame of the speech signal while at the
same time respecting the energy transitions from frame
to frame without thereby affecting the computation load
and the necessary memory space in the vocoder.
BRIEF DESCRIPTION OF THE DRAWINGS
Other features and advantages of the invention
shall appear~from~the~followlng description, made with
reference to the appended drawings, of which:
:
-- 2~833~
s
.
- Figures 1 and 2 show two graphs to illustrate
the principle of quantification of the energy of a
vocoder implemented by the inventlon;
- Figure 3 is a flow chart illus~rating the
different steps of the method according to the
invention.
MOEE DETAILED DESCRIPTIOM~ ~
The method~according to the lnvention consists, in
:
the manne~ shown in fi~ure 1, in segmenting the speech
signal into irames wlth a constant determlned duration
ranging, for examplel from 22.5 to 30 ms, grouping the
frames in packets of a determined number n of energy
values of the signal in each frame to transmit, in each
packet, only~the first quantified value of the energy
~measured~El in the~first frame of a packet as well as
the k - l values o~ the diffPrences of the energies
existing between the frames that follow, k being equal
to n.L. In reception, the differences of the energies
received are placed end to end after the first energy
value that~;is received in the first frame of each
packet to reconstitute the profile of the quantified
values of the energies at emission.
To do this, in the emission vocoder, a first value
:: :
; of energy is quantified in each first frame ko of a
packet in a determined number QO of bits and the
:: ~ : ::
variations of ~ the k~ - 1 remaining energies are
quantified with a determined number Q1 of bits smaller
. . . : . . :
: - ~ .
'
2~3~3~
than QO. The 2Qo possible initial values include a zero
value representing the silences. ThP other values are
distributed according to an almost logarithmic scale
which is best suited to following the properties of
sensitivity of the ear: the higher the level of the
speech signal, the smaller is the quantification step.
Typically, a 3dB step is adopted for the low levels and
a 1 dB step is adopted for the high levels. The m = 2Q1
other values represent energy increments d; also
referred to hereinafter as "legal values of energy",
the values of which are predetermined to emphasize the
transitions. These transitions are chosen for example
as being respectively e~ual to -3dR, OdB, +2dB and +7dB
f the number Ql ls~encoded with only two bits.
As can be~seen ln figure 2, the energy increments
can be used to make a search, from each quantified
value B of a frame k, for the quantified values ~ of
the energy in the k - 1th preceding ~rame which could
lead to said value B by a legal increment dj starting
with the zero increment do~
The numbers QO and Q1 are determined according to
the steps 1 to S of the method represented by the flow
chart of figure; 3. The flrst step referenced 1 in
figure 3 groups together the frames in packets of L
frames. The values of the energies E1 to Ek are
computed~at the step 2. These are quantified in the
manner shown in figures 1 and 2 between two values Emax
- ~$3~3~
and Emjn in relation to a scale comprising P
graduatlons which may be identified for convenience's
sake with the 2Qo possible values of the initial energy
E1 measured in the first frame. The quantified values
corresponding to the 2Qo posslble values are designated
in figure 2 by eO, e1 ... ep_1 with eO = Emjn and ep_1
= Emax
The method continues at the step 3 in figure 3 by
an initialization stage in which a set of P distances
is computed between the first value of energy E and the
P possible quantified value.s of this energy.
The corresponding distances Dp are memorized in
the form of a first table (D~, not shown, in a memory
of the vocoder. The computations take place by squaring
the differences between the first energy EI and the
quantified values eO, el... ep_1 according to the
relationship:
D(p) = (El - ep)2 where p = 0, 1 ... P - 1
The computed distances are all the smaller as the
quantified value ep is closer to the value El. The next
step 4 consists, in a manner similar to the known
VITERBI algorit~mt in in carrying out k - 1 iterations
aimed at estimating the distances between all the
potential quantification profiles and the real energy
profile, in eliminating the least probable
quantification profiles. ~ second table (D') not shown
and referenced ~Islope~ is prepared. For each of the
- . . .
. - ., -
.
. .
- 2~33~
iterations l to k - l r this second table D' associates
a slope or a legal energy increment dj with each
quantified value P of the i.teration k. A search for the
quantified value of the preceding k - 1th iteration is
; 5 done by the ticking off, in the "slopes" table, of the
"part" or legal increment dj that can lead directly
thereto, beginning with the zero increment do. The
sequence ~of ~the programming instructions to be
implemented is the following:
- FOR p = 0 ... P - l, DO
/* initiaIization for a zero incrementation*/
- Let Dm;n = D (p~do~ = D' (p) and let PrecIndex = O
/*test of:the~non-zero Lncrementations */
- FOR i =~1 ... m, DO
~: - If p - dj > = O AND p - dj < = P - 1 THEN /*legal value dl*/
- If D'(p - dj) < Dmin then /* shorter distance */
DO Dmjn = ~' (P dj)
- DO PrecIndex = i
- END IF
- END IF
END DO
~ DO SlopeIndex ~k~p)=precIndex/* memoriæe the most
probable quant1fied value at the preceding step*/
- DO D(p) = Dmjn -~ Ek-ep)2/* update the distance*/
END DO
.. . . . .
':' ''. ' ' ' : `' ` ` ~
,
.
` ` ` . '
2~3~3~
Thus, at -the k - 1th iteration, a table of
distances D~) is prepared. This table, at the position
p, contains the cumulated distance between the best
quantified profile that arrives at the position p and
the original profile. This makes it possible to keep,
in memory, a table of slope indices wherein the slope
index value (k, p) represents the index of the best
possible s~ope to arrive at the quantified value ep at
the step k. The two tables thus obtained make it
1~ possible~to arri.ve at a fina} decision. To do this, the
method entails carrying out a search in the table D(+)
for the index Pmin which corresponds to t~e minimum
value. Then it conslsts in making a trace-back in the
slopes table by carrying out k - 1 iterations
programmed as follows:
- for k = K - 1, K - 2, ...., 1 DO
- Dif~Index(f) = SlopeIndex(k,p
~ Pmi n = Pmi n - SlopeIndex(kl Pmi n )
END DO
The index values Index Diff (1 .......... K - 1) are the
indices of the best quantified values possible for the
slopes Dj. The final value of Pmin is then simply the
most probable quantified value~
The correspondence between the original profile o~
the values o~ the energies to be quantified after the
final profile~after quantification is shown in figure
1. The fact that the algorithm automatically eliminates
' ~ . ': : . - ' ' .
- : - ~ ., ,: ' ,... . .
' ' ''' '. .
. ,' :~, ~ ,
- - 2~8~33~
the aberrant values resulting from a false analysis
appears in the fourth value of energy shown in figure
1.
Naturally, the method that has just been described
can always be matched to particular characteristics of
the system of analysis. In particular, 1f this system
tends to find erroneous values for energy, it is always
possible to minimize the influence of the erroneous
values through the replacement, for example, of the
squaring operations used for the distance measurements
by absolute values that enable the profile of the
quantified values to be linked with the correct values
of energy, provided that they are more numerous than
the incorrect~values.
Furthermore, the operat1ons of matching and fine
tuning fcr the vocoder require only modifications of
the quantified starting values (number and values), the
increments (number and values), or again the number of
iterations.
~ Finally, the method that has just been described
represents only a small computation load since the
initialization is done starting with the very first
frame, and the kth iteration is done at the k ~ 1
frame. This.enables the distribution of the computation
load in time, except for the last frame where the final
decision is taken without the arrangement's being
costly in terms of computation power.
'
