 |
Description  |
|
|
BACKGROUND OF THE INVENTION
1. Field of the Invention
This present invention relates generally to the field of data-compression
of digitally encoded speech signals and more particularly to a system and
a method for compressing the serial bit stream signals that are generated
by delta modulation encoders, such as CVSD (continuous variable slope
modulation) encoders, and by sampled clipped speech encoders
2. Description of Related Art
A patent of interest for its teaching of analog to digital conversion of
speech signals is U.S. Pat. No. 4,271,332 entitled "Speech Signal A/D
Converter Using an Instantaneously-Variable Band Width Filter", by J. C.
Anderson.
The signal technique described in this patent is denoted the MIMIC
technique which will be referred to in the Description of The Preferred
Embodiment
A patent of interest for its teaching of processing sampled clipped speech
signals is U.S. Pat. No. 4,594,575 entitled "Improved Digital Processor
for Speech Signals" by Avery et al. The type of processor described in
this patent will be denoted SPFE, an acronym for Speech Processor
Front-End. The acronym will be used in the Description of The Preferred
Embodiment.
Machines that talk have been popular for many years, for they take on
almost human characteristics in questioning and response.
Although many speech-synthesis products exists, their characteristics
differ widely and are influenced by many factors such as speech-encoding
methods, bit rate, and vocabulary preparation.
From the different sound and tonal inflections that accompany the spoken
word, it is obvious that speech properties change during the transition
from voiced to unvoiced speech. For example, there are large changes in
peak amplitude and fundamental frequency, but because waveforms change
little for short segments of speech, most speech-processing techniques,
which intend to achieve a low bit rate, isolate such segments and process
them as if they were short segments of sustained sound having fixed
properties. This segment action which is usually referred to as framing,
introduces a distortion in the reconstruction of speech. This distortion
degrades the speech quality and intelligibility. For this reason and
others, most of the low bit rate speech processors sound mechanical.
The need for deciding whether a given segment of a speech waveform should
be classified as voiced speech, unvoiced speech, or silence (absence of
speech) arises in many speech analysis systems. Most of the available
methods work in conjunction with pitch analysis to decide what class the
segment should fall into. There are two disadvantages in this approach to
the voiced-unvoiced decision. First, the decision is based on a single
feature--the degree of voice periodicity. Voiced speech is only
approximately periodic; sudden changes in articulation and the
idiosyncrasies of vocal cord vibrations can produce speech waveforms which
are not periodic. In such cases, a feature such as the amplitude of the
largest cepstral peak will fail to distinguish voiced speech from
unvoiced. Second, the voiced-unvoiced decision is tied to the pitch
detection which may be acceptable for speech synthesis applications, but
for the proposed application, the linking of the voiced- unvoiced decision
to pitch detection can result in unnecessary complexity as well as in
poorer performance, particularly at the boundaries between voiced and
unvoiced speech.
Pitch (i.e., fundamental frequency F.sub.0 and fundamental period T.sub.0)
occupies a key position in the acoustic speech signal. The prosodic
information of an utterance is predominantly determined by this parameter.
The ear is more sensitive to changes of fundamental frequency than to
changes of other speech signal parameters by an order of magnitude. The
quality of voiced speech is essential influenced by the quality and
faultlessness of the pitch measurement. The importance of this parameter
thus necessitates using a good and reliable measurement method. The prior
art shows that pitch detection can roughly be divided into the following
three broad categories:
(1) A group which utilizes principally the time-domain properties of speech
signals;
(2) A group which utilizes principally the frequency-domain properties of
speech signals; and
(3) A group which utilizes both the time and frequency domain properties of
speech signals.
SUMMARY OF THE INVENTION
The present compression system adopts a method which uses a pattern
recognition approach for classifying a given speech segment into three
classes: voiced speech, unvoiced speech and silence. The pattern
recognition approach provides an effective method of combining the
contributions of a number of speech measurements, which individually may
not be sufficient to discriminate between the classes, into a single
measure capable of providing reliable separation between the three
classes. The system and method implemented herein is essentially based on
the rate of the zero crossings of the speech waveform from the bit serial
encoder, namely CVSD, MIMIC, or SPFE, used for digitizing the speech
waveforms.
The following three measurements used in the system and method of the
present invention are:
(1) The number of silent patterns (0000, 0101 or 1010) of SPFE, MIMIC, and
CVSD respectively "X0".
(2 ) The number of one pulses in a-frame "S1".
(3) The number of one bits in a frame "S".
(4) The width of the average one pulse "S/S1".
The choice of these particular parameters is based partly on experimental
evidence that the parameters vary consistently from one class to another
and partly on the knowledge of the method in which voiced and unvoiced
speech sounds are generated in the human vocal tract and how each of the
bit serial encoders digitizes these speech sounds- The present compression
system and method takes into account the changes in the speech properties
during the transition from voiced to unvoiced speech and avoids
segmentation in order to preserve the quality and intelligibility of the
reconstructed speech. Although it was mentioned earlier that the adopted
method is based partly on the rate of zero crossings of the speech
waveform, from an analog point of view, it is appreciated that since the
digital output of a bit serial encoder is used for classifying speech in
the three classes of sounds, that the rate of the zero crossing of the
speech waveform corresponds to the rate of change from a high to a low
(one to zero) or vice versa in the serial bit stream of the encoder. On
the other hand, the use of the digital bit stream for reliably
discriminating between the three classes of sounds across a wide variety
of speakers simplifies the detection mechanism. The present system and
method has no need to measure the energy of the signal, largest cepstral
peak, and variation of the zero-crossing count from one speaking
environment to another which reflects the variable characteristics of the
room noise, because the serial encoders are designed to take care of these
aspects.
The pitch detection portion of the present speech compression system and
method determines the pitch during a voiced frame of speech and relies on
the time and frequency domain properties of the digital stream of the
encoders. For this pitch detection, the measurements are zero-crossing
measurements and autocorrelation measurements. The zero-crossing
measurements are made by locating the large one pulses in the bit stream
of the encoder, while the autocorrelation measurement is made by counting
the number of one bits in a frame. The basic assumption in estimating the
pitch period is that if the speech waveform is periodic in the time
domain, then the digital bit stream of the encoders will consist of a
series of large one pulses at the fundamental frequency and its harmonics.
Thus, simple measurement can be made on the bit stream of a frame to
estimate the pitch period. A voiced frame outputs a large one pulse at the
fundamental frequency of the speech waveform. The pitch detector portion
places pitch markers directly on the bit stream of each frame. To obtain
the pitch markers of interest for each voiced frame, the pitch detector
portion calculates the average width of one pulses, the location of the
one pulses within the frame, the width of every one pulse, and finally the
number of these one pulses. From these measurements, the pitch detector
portion selects the one pulses whose width is greater than the average one
pulse and accordingly saves the location of these large one pulses within
the frame. The remainder of the system isolates and identifies the
principal excursion cycles, i.e., those which correspond to true pitch
periods. This is accomplished through a series of steps by using the range
of the fundamental frequency observed in human speech and the constraint
on the frame size which force the frame to have at least two pitch
periods. The excursion cycles are further identified by justifying that
the distance between two consecutive pitch markers, which is the estimated
pitch period, is within the fundamental frequency range of human speech
and does not exceed half the frame size. The pitch detector portion sorts
three estimated pitch periods in ascending order and then divides them
into groups. Finally, the smallest element of the largest group is the
pitch period that is selected. For a silent frame, the pitch period is set
equal to 8 in the preferred embodiment of the invention, because of the
repetitive byte recognized in the bit stream of the bit serial encoders.
For an unvoiced frame, the pitch period is set equal to the frame size due
to the lack of pitch markers.
One of the unique properties of the pitch detector portion of the present
invention is the usage of the digital bit stream in estimating the pitch
period rather than the speech waveform. This eliminates several problems
encountered by pitch detectors which use the analog speech waveform to
estimate the pitch period.
The operation of the preferred embodiment will be described in conjunction
with the digitally encoded word "STAY" but it will be obvious to those
persons skilled in the art that this description is for example only and
is not to be taken by way of limitation as to the capabilities of the
present invention.
Briefly the method of the present invention is as follows: a serial bit
stream of speech encoded data is stored in a temporary buffer storage
before compression. A microprocessor is adapted to read the stored data in
fixed length frames. Each frame is processed to classify whether the frame
represents voiced, unvoiced, or a silence condition. For a voiced frame,
the microprocessor detects a pitch. The bits of this voiced frame, are
then compared with the bits of prior received voiced frames, on a pitch
basis, so as to detect repetition. When repetition is detected, the
repetition count, the pitch period, and the repetitive pitch bits are
stored in another temporary storage, thereby compressing the number of
bits that have to be stored and/or transmitted. The compressed information
is reconstructed by utilizing a decoder which can interpret the stored
information and cause reinsertion of the repetitions.
Accordingly, it is a primary object of the present invention to provide an
improved system and method for compressing speech signals.
It is another object of the present invention to provide a system and a
method which considers the changes in speech properties during the
transition from voiced to unvoiced speech.
It is yet another object of the present invention to unlink the
voiced-unvoiced decision from the pitch detection.
It is a further object of the present invention to provide a system and a
method which operates upon the zero-crossings of the speech waveform from
bit serial encoders.
These and other objects of the present invention will become more apparent
and better understood when taken in conjunction with the following
description and the accompanying drawings, through which like characters
indicate like parts and which drawings form a part of this application.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 illustrates in system diagram form the preferred embodiment of the
present invention.
FIG. 2 illustrates the preferred embodiment of the invention in block
diagram form.
FIG. 3 is a detailed block diagram illustrating a processor encoder portion
of the system of FIG. 1.
FIG. 4 is a flow chart illustrating the operation of a voiced/unvoiced
detector subassembly of FIG. 3.
FIGS. 5A and 5B, arranged according to the map of FIG. 5, illustrate a flow
chart depicting the operation of a pitch detector subassembly of FIG. 3.
FIG. 6 illustrates various types of compressed data blocks.
FIGS. 7A-7C, arranged according to the map of FIG. 7, illustrate a flow
chart corresponding to the pitch interval repetition algorithm of FIG. 3.
FIGS. 8A and 8B are a detailed block diagram and a flow chart,
respectively, illustrating the processor decoder portion of the system of
FIG. 2.
FIGS. 9A-9D are waveforms illustrating the word "STAY" in analog form,
SPFE, CVSD, and MIMIC digital form, respectively.
FIGS. 10A-10C are waveform representations of the spoken word "STAY"
processed through a SPFE encoder.
FIGS. 11A-11C are waveform representations of the spoken word "STAY"
processed through a CVSD encoder.
FIGS. 12A-12C are waveform representations of the spoken word
"STAY"processed through a MIMIC encoder.
FIGS. 13A-13C are waveform representations illustrating pitch detection of
the spoken word "STAY" for SPFE, CVSD, and MIMIC encoded voice frames,
respectively.
DESCRIPTION OF THE PREFERRED EMBODIMENT
Referring to FIG. 1, a source of electrical speech signals 10, which may,
for example, be a tape recorder or a microphone, is connected to deliver
electrical analog speech signals to the input of a bit serial encoder
(digitizer) 20. The digitizer 20 can be of a type that provides a serial
bit stream using either a SPFE, CVSD, or MIMIC based system. The serial
bit stream from the digitizer output is stored in a temporary storage 30.
The serial bit stream may be stored on a floppy disc 32 which disc is
transferred to the disk reader 34 portion of a processor 40 when read from
storage is required. The processor 40 which may be an NCR Decision Mate V,
operates upon the stored data to compress it, temporarily store it, and
transmit it to a bit serial decoder 60. The compressed information is
reconstructed into the original analog signal by the bit serial decoder
60. The analog signal from the decoder 60 is fed to an "Amplifier-Speaker
unit" 70 or other applicable device for generating the reconstructed
analog speech.
The processor 40 reads the original data from the temporary storage 30 in
frames and processes the original data with a compression algorithm. Each
of these frames have a fixed length, IW bits long. Experimentally it was
determined that the frame length is dependent on the bit rate of the
digitizer and should be at least two pitch periods wide where a pitch
period ranges from 80 to 300 Hz.
Referring to FIG. 2, which is the block equivalent of the elements shown in
FIG. 1; the processor 40 is comprised of a receiver 41 , a compression
algorithm 42, temporary storage 50, decoding algorithm 43 and transmitter
44. The receiver 41 reads the original data stored by the temporary
storage subassembly 30 in fixed length and passes these frames to the
compression algorithm 42. When reconstruction is desired the processor
decodes the stored compressed data and transmits the decoded data as a
serial bit stream to the bit serial decoder 60. Each frame is processed as
an array IB[] by the compression algorithm 42 (shown in block form in FIG.
3). IB is an IW .times.1 array. The process flow of the compression
algorithm 42 is set out in FIGS. 4, 5A, 5B, 7A, 7B, and 7C. The operation
of the compression algorithm will be described later in this section of
the specification. The output of the compression algorithm is a sequence
of compressed blocks, the form of which is illustrated in FIG. 6, which
are stored in the temporary storage 50. Upon readout a decoding algorithm
43 (shown in block form in FIG. 8A provides a repetition reinsertion to
the compressed data. The transmitter section 44 transmits the data from
the decoding algorithm to the bit serial decoder 60. The process flow of
the decoding algorithm 43 is set out in FIG. 8B.
Referring to FIG. 3, original data received by the receiver 41 is formed
into fixed length frames IB[] by a fixed length framing element 45. Each
frame array IB[] is directed to a voiced/unvoiced/silence (V/UNV/S)
detector 46, a pitch detector 47 and a pitch interval repetition (P.I.R.)
algorithm 48. The V/UNV/S detector 46 functions to class each frame array
IB[] by calculating a value X.sub.0, where X.sub.0 is equal to the number
of consecutive sequences which have the patterns "0000", "0101" or "1010".
These patterns have been found to repeat indefinitely in a silent frame.
X.sub.0 is compared against threshold values to determine the class of the
frame. A silent frame is defined as CLASS 0. An unvoiced frame is defined
as CLASS 1, and a voiced frame is defined as CLASS 2. The class
determination is directed as inputs to the pitch detector 47 and to the
pitch interval repetition algorithm 48. The pitch detector 47 uses the
array IB[] to detect pitch, based on what class the frame is. A more
detailed description of the operation of the pitch detector will be made
with the description of the flow of FIGS. 5A and 5B. The output of the
pitch detector 47 is denoted IP and is directed as an input to the pitch
interval repetition algorithm 48. The P.I.R. functions to calculate a
value P which equals the number of pitch periods in a frame. The P.I.R.
provides as an output a frame residue signal (if any) in the frame IB[]
that has not yet been compared. The residue signal is directed back as an
input to the fixed length framing element 45. A FLAG signal, which is
equal to 2 if the previous processed frame was voiced and is equal to 0 if
it was not voiced, is also directed back as an input to the fixed length
framing element 45. The P.I.R. outputs a sequence of compressed data
blocks (See FIG. 6) for storage in the temporary storage for compressed
data 50.
Referring to FIG. 4 the voiced/unvoiced silence detector 46 operates upon
the array IB[] with a flow that commences with the START block 101. In
block 102 the value X.sub.0 is calculated for each IB[]. X.sub.0 is equal
to the number of "0000", "0101", or "1010" sequences in a frame. When
X.sub.0 is computed, a decision block 103 determines if X.sub.0 is greater
than or equal to a value X .sub.2. If the answer is "yes" a block 104 sets
the CLASS signal equal to 0. If the answer is "no" a block 105 operates to
calculate S and S1 which are values representing the number of one bits
and the number of one pulses, respectively, in an array IB[]. A one pulse
is defined as, a pulse which has one or more consecutive one bits. A "no"
decision from block 103 indicates that a frame may be voiced or unvoiced.
If the ratio of S/S1 (average width of one pulse) is less than or equal to
X.sub.1, then the frame is unvoiced and the CLASS is set equal to 1 by
block 107 otherwise the frame is voiced and the CLASS is set equal to 2 by
block 108. The value X.sub.1 is an unvoiced threshold level and the value
X.sub.2 is a silence threshold level determined empirically by examining a
number of unvoiced and silence frame patterns, respectively. The flow then
ends at the END block 109 with the determined CLASS signal being directed
to the pitch detector 47 and the P.I.R. algorithm 48.
Referring to the flow of FIGS. 5A and 5B, which relates to the operation of
the pitch detector 47, the flow commences at the START block 201 and moves
to the decision block 202. A "yes" decision indicates that the frame is
silence, CLASS=0 and a decision block 203 is entered to determine if the
previous frame was voiced by noting if the FLAG is equal to 2. If the
previous frame was voiced the previous pitch is left unchanged and the
flow proceeds to the END block 210. If the FLAG does not equal 2 then a
default pitch IP is set equal to 8 in action block 204. Referring back to
the decision block 202, if the decision is "no" the CLASS is not equal to
"0" and the decision block 206 is entered to determine if the CLASS is
equal to 1. If the decision is "yes" a further check is made by decision
block 207 to determine if the FLAG is equal to 2. If "yes", the algorithm
does not change the previous pitch. If the decision is "no" (i.e., FLAG
not equal to 2), the pitch IP is set equal to a frame size IW (number of
the bits in a frame) by the action block 208 and is directed to the END
block 210. Referring back to the decision block 206, if the frame is
voiced, that is the CLASS=2, the algorithm calculates the average width of
a one pulse X6 within the action block 209, and then the location of each
one pulse in the frame IB[] in action block 210. An array P1[I] is then
calculated in action block 211 which corresponds to the width of the one
pulses. These widths are saved in the array P1[I]. At action block 212, by
using the arrays P1[I], W[I], and X6, the algorithm selects only the large
one pulses, and saves their locations in an array P2[J] and their widths
in an array P3[] at action block 213. A large one pulse is a pulse whose
width is greater than or equal to X6. By using P2[J] and P3[], the
distances between the large zero pulses is calculated and saved in an
array P[I] at action block 218. By using the array P[I] at action block
219, the algorithm obtains the array Q[] of pitch period estimates such
that each estimate is equal to the distance between the large one pulses
in each pair of large one pulses in the frame satisfying the following
constraints:
(a) The distance between the large pulses in each pair is greater than or
equal to X3 and less than or equal to IW/2. X3 is a distance threshold
obtained empirically.
(b) The widths of the two pulses in the pair are approximately the same,
within a tolerance of .+-.X4. X4 is a width threshold obtained
empirically.
After the array Q[] is found, the algorithm sorts the entries of Q[] in
ascending order at action block 214. The sorted Q[] array is divided into
the defined classes at action block 215. Each class consists of values
that are within X5 from each other. X5 is also an empirically determined
threshold. The algorithm then selects the biggest class (i.e., the class
that has the maximum number of elements) at action block 216 and picks the
minimum value of this class as a pitch period estimate (IP) at action
block 217. The minimum value is then directed to the END block 210 to end
this portion of the flow.
The frame CLASS and the pitch estimate IP are then fed to the P.I.R.
algorithm, the flow of which is depicted in FIGS. 7A, 7B and 7C. This flow
commences at the START block 300. A decision block 301 determines whether
the frame is voiced (i.e., CLASS=2) or not. If "yes", a determination is
made, at decision block 302, as to whether the previous frame was voiced
(i.e., the FLAG=2) or not. If "no", action block 303 sets the FLAG =2 and
at action block 304, N is set equal to 1 and the flow moves to action
block 305 wherein the value P is calculated. The value N is a repetition
count corresponding to the number of times that data will be repeated on
reconstruction and the value P equals the number of pitch periods in the
frame being processed. An array RP[], corresponding to the first IP bits
of the frame IB[], is chosen by action block 306 as the reference pitch
period pattern. At action block 307 the RP[] pattern is compared with the
bit pattern in the next group of IP bits in IB[]. The comparison is done
on a bit-by-bit basis in a continuous order starting from the first bit in
RP[] and the first bit in the group being compared. A value, SS, is set
equal to the number of bits that do not match in the previous comparison.
The value N is checked at decision block 308 to determine whether it is
greater than or equal to 255. If "yes" (i.e., N >=255), the algorithm at
action block 311 stores the values of N, IP, and RP[] in the temporary
storage 50 in compressed data format as shown in FIG. 6. Then, at action
block 312, N is set equal to 1 and at action block 313 the RP[] pattern is
set equal to the last compared group of IP bits in IB[]. If N is less than
255, the algorithm at decision block 309 compares whether SS is less than
or equal to XN, where XN is an acceptable matching bit thresholds and if
the answer is "no", then the same sequence of events, as when N is greater
than or equal to 255, is executed. If the answer is "yes" that is SS
<=XN, the algorithm at action block 310 increments N by 1 and directs the
flow to decision block 314. At decision block 314, if there are more IP
bits in IB[] to be compared, the algorithm flows back to action block 307
to compare the current RP[] with a group of IP bits of IB[] that have not
yet been compared and continues the same checking as described above. If
the decision is "no" (i.e., no more IP bits left in IB[]), then IP1[], at
action block 315, is set equal to the frame residue, that is, to the
remaining bit sequence (if any) in IB[] not yet compared. Finally the
algorithm goes back to the fixed length framer 45. Returning to the
decision block 302, if a class is voiced and the previous frame is voiced
(i.e., CLASS=2 and FLAG=2), then the algorithm jumps to action. block 307
for the comparison of RP[] and the next group of IP bits in IB[] and the
same flow of events (as before) are executed thereon.
Referring back to decision block 301, if the frame is not voiced (i.e.,
CLASS is not equal to 2), the flow is directed to decision block 316 to
determine if the frame is unvoiced (i.e., CLASS=1) or silence (i.e.,
CLASS=0). If the frame is unvoiced, the flow is directed to a decision
block 317 to determine if the previous frame was voiced (i.e., FLAG=2). If
"yes", the algorithm stores N, IP, and RP[] in the temporary storage 50,
in compressed data format, at action block 318, and sets, IP equal to IW
at action block 321. IW equals the number of bits in a frame. The FLAG is
set equal to 0 at action block 320 and is set equal to 1 at action block
319. If the decision, at decision block 317, is "no", (FLAG not equal to
2), then the main unvoiced stream of processing starts at action block 319
by setting N equal to 1 and proceeds to storing N, IP, and IB[] in the
temporary storage 50 in compressed data format at action block 322. If the
frame is not unvoiced (i.e., CLASS not equal to 1), then the frame must be
silence (i.e., CLASS=0) as determined by decision block 316. For a silence
frame, the algorithm checks at decision block 323 to determine whether the
previous frame is voiced (i.e., FLAG=2 ), if "yes", it stores, at action
block 324, the values of N, IP, RP[] in the temporary storage 50, in
compressed data format, and sets IP equal to 8 at action block 325 and the
FLAG equal to 0 at action block 326. The flow then merges to the main
silent stream of processing at action block 327. If the decision, at
decision block 323 is "no", (i.e., FLAG not equal to 2 ) then the main
silent processing flow starts by setting IB[i] equal to bi for i=1, 2, . .
. , 8. The pattern, b.sub.8 b.sub.7 b.sub.6 b.sub.5 b.sub.4 b.sub.3
b.sub.2 b.sub.l, was determined empirically to repeat itself indefinitely
for the entire silent frame. After IB[i] is set, the algorithm sets N
equal to the integer part of the ratio IW/8 so that the pattern, b.sub.8
b.sub.7 b.sub.6 b.sub.5 b.sub.4 b.sub.3 b.sub.2 b.sub.l, is repeated N
times. The flow then moves to decision block 328 to determine if C is
greater than 2. If "no", C is incremented by 1 at action block 330 and the
flow is directed to action block 322 for storing the value of N, IP and
IB[] in the temporary storage 50 in compressed data format, and from there
the flow goes back to the framer 45. If the decision at decision block 328
is "yes", (i.e., C<=2), the algorithm sets C equal to 0 at action block
329 and then goes to the framer 45. The value C is a count of the
consecutive silent frames and is set equal to 1 when the first frame is
read. This count does not allow the algorithm to store more than three
consecutive silent frames. The silent criteria of the pitch interval
repetition is added to the algorithm to remove pauses between words or
sentences.
Referring now to FIG. 8A, the decoder reads the compressed data from the
temporary storage 50 in the order they were stored. First, in block 51,
the repetition count N is read from the first byte of the bits. Second, in
block 51, IP in the data field, is read from the second byte. Thirdly, the
data fields are read, at the data field reader 52, from the third byte up
till the nth byte. This information is used at block 53 to repeatedly
reinsert the IP bits in the data field N times resulting in an expanded
data format. This expanded data format or serial bit stream is fed to an
output drive 54 which is part of the transmitter 44 for transmitting the
bit stream to the bit serial decoder 60 for playback. This process of
repetition, reinsertion, transmission, and playback of the compressed data
is repeated until all of the compressed data, stored in the temporary
storage 50, is processed.
Referring to FIG. 8B, the compressed data blocks from the temporary storage
50 are read out, one at a time, with action block 401 reading the value of
N from first byte of compressed data. The action block 402 then reads the
value IP from the second byte of compressed data. The number of bytes in
the data field BN is determined in action block 403. Counter variables I
and J are set equal to 1 and 0, respectively, in action block 404. The
action block 405, decision block 407, and count incrementer block 406
operate as a "D0-loop" which continues to repeat the action of action
block 405 until I is greater than BN. When that condition is reached, an L
counter block 408 is set equal to 1 and, action block 409 stores the bytes
of the array IB[M] in a decompressed file. The value L is checked at
decision block 410 to determine whether it is greater than or equal to N
(repetition count). The L counter 411 is incremented by 1 for each "no"
answer from decision block 410. If the decision is "yes" a check is made,
at decision block 412, to determine if there are any more compressed data
blocks in the temporary storage 50. If the answer is "yes", action block
401 starts reading the next compressed data block. If the answer is "no",
the decompressed data is directed to the output drive 44 and the bit
serial decoder 60. In the preferred embodiment of the invention the output
drive 44 is coupled to the bit serial decoder 60 by a Centronics parallel
printer I/0 port (not shown for simplicity). The output drive reads the
file generated by the repetition reinsertion block 409 and transfers the
data to the bit serial decoder 60.
TABLE 1
______________________________________
Experimental Thresholds
X1 X3 X4 X5
Unvoiced X2 Distance
Width Class
Thresh- Silence Thresh-
Thresh-
Thresh-
Encoders
old Threshold old old old
______________________________________
CVSD R/4.0 0.225 (IW)
R/300 1 8
MIMIC R/4.8 0.225 (IW)
R/300 1 5
SPFE R/4.8 0.225 (IW)
R/300 1 5
______________________________________
The above table sets forth the experimental threshold levels for three
different digitizers that can be used in this compression system. Note
that XN (matching threshold) was not included in the table because XN can
be set equal to 10, 15 or 20 depending on the compression factor needed
and the speech quality desired by the user.
DETERMINATION OF THE EXPERIMENTAL THRESHOLDS FOR THE SILENCE DETECTOR AND
PITCH DETECTOR
Referring now to FIG. 9A, the analog signal representation of the spoken
word "STAY" is shown divided into its unvoiced, silence and voiced
portions. The voiced portion is seen to contain the largest amplitude
variations as compared to the other portions.
FIGS. 9B, 9C and 9D illustrate the serial bit stream appearing at the
output of a SPFE, CVSD and MIMIC encoder, respectively when the signal of
FIG. 9A is applied to their inputs.
The voiced/ unvoiced/ silence detector 46 and the pitch detector 47
parameters were determined by analyzing the bit pattern of several speech
waveforms which were digitized by different encoders, namely SPFE, CVSD,
and MIMIC. The analysis was conducted for different speakers and bit rates
(9.6, 16 and 40 Kbit/sec). The pattern analysis was performed by using a
digital Sona-Graph for displaying the speech waveforms and the
corresponding bit patterns of these speech signals, and measuring the
following parameters:
(1) Experimental unvoiced threshold "Xl":
After investigating the bit pattern of voiced frames versus unvoiced
frames, it was found that the serial bit streams of each encoder during an
unvoiced frame changes from high to low (i.e one to zero or vice versa)
more rapidly than the serial bit stream of a voiced frame. This criteria
of unvoiced frames imposed a restriction on the width of the one pulses.
From this restriction the unvoiced threshold "X.sub.1 " was determined as
a powerful parameter for discriminating between a voiced frame and an
unvoiced frame. This characteristic of unvoiced frames is justified by the
fact that unvoiced sounds are generated when the vocal cords are not
vibrating and its energy spectrum looks like the spectrum of broadband
white noise. These criteria cause several zero crossing in the speech
waveform of an unvoiced frame. Consequently, the digital output of SPFE,
CVSD, and MIMIC will consist of short duration one pulses.
FIGS. 10A-10C, 11A-11C and 12A-12C, show typical silent, unvoiced, and
voiced frames for SPFE, CVSD, and MIMIC respectively. These figures
represent a greatly expanded view of the signals shown in FIGS. 9A to 9D.
The 1 labeled figures correspond to the like labeled analog portion of
FIG. 9A, while the 2 labeled figures correspond to the like related
digital portions of the encoder serial bit stream. Referring now
specifically to FIGS. 10A-10C, 11A-11C and 12A-12C, it is obvious that the
bit pattern of the unvoiced frame can be discriminated from that of a
voiced frame by using X.sub.1. The following table lists the variations of
X.sub.1 for different speakers and different bit rates.
TABLE 2
______________________________________
Time Duration Range
Speakers: of X.sub.1
______________________________________
F1: Female (No. 1) 0.10 to 0.20 msec
F2: Female (No. 2) 0.104 to 0.205 msec
M1: Male (No. 1) 0.10 to 0.201 msec
M2: Male (No. 2) 0.103 to 0.208 msec
______________________________________
Width of X1 in
Width of X1 in
Width of X1 in
Bits at 9.6 Kbps
Bits at 16 Kbps
Bits at 40 Kbps
______________________________________
F1: 1 to 2 bits 2 to 4 bits 4 to 8 bits
F2: 1 to 2 bits 2 to 4 bits 4 to 8 bits
M1: 1 to 2 bits 2 to 4 bits 4 to 8 bits
M2: 1 to 2 bits 2 to 4 bits 4 to 8 bits
______________________________________
(2) Silence Threshold level "X2" | | |
