 |
Description  |
|
|
FIELD OF THE INVENTION
This invention relates to digital to analog and analog to digital
converters in general and more specifically to analog to digital speech
sampling and encoding means for preparing one or more speech signal
sources for transmission in a digital link for reconstruction by digital
to analog converters at the receiver. It is particularly suited to
multiple voice channel compression for transmission.
BACKGROUND OF THE INVENTION
Numerous sub-band speech coders are known as are speech coder systems based
on digital microprocessors for handling the manipulation of the digital
energy level samples normally occurring in such systems. Reference may be
had to the 1977 IEEE International Conference on Acoustics, Speech and
Signal Processing record of May 9-11, 1977, pages 191-195 and to the IEEE
Acoustics, Speech and Signals Processing Society Proceedings of April
9-11, 1980, Vol. 1, pages 332-335 which show typical digital sub-band
speech coder and decoder arrangements for multi-channel speech
transmission use.
In systems like those in the above-referenced publications, the subjective
quality of performance of sub-band speech coders perceived by a listener
at the receiver is highly dependent upon the allocation of available bits
in the transmission medium to the individual frequency bands in the
sub-band coder. A major improvement previously made involved the use of
dynamic bit allocation where the available bits are dynamically
distributed among the frequency bands according to the energy present in
each band sample. This technique was extended to a variable bit rate
system where many speech coders could share a common bit rate resource,
i.e., a transmission channel, by assigning bits to all the bands of all
the coders according to the energy in each frequency band in its
relationship to all other frequency bands.
The typical sub-band speech coder takes the 0 to 4 kilohertz speech
spectrum and samples it, typically at a sampling rate of 8,000 samples per
second. Through filtering and sub-sampling, the speech spectrum is divided
into sub-spectra, typically into eight sub-bands of 500 Hertz width each.
In such a system, depicted schematically in FIG. 1, incoming analog
signals on analog line 1 are converted to digital sample stream by the
analog to digital converter 2, samples of which are clocked out by the
clock 4 over line 3, typically at an 8 kilohertz sampling rate to a
parallel filter bank 5.
The filter bank 5 divides the incoming digital stream into typically 8
frequency sub-bands spanning the spectrum from 0 to 4000 hertz. The output
is thus a series of eight individual channels each having samples
occurring at the rate of 1000 samples per second as schematically shown by
the clock 6 controlling the output of the filter bank 5 over lines 7.
Individual frequency sub-band peaks and the overall peaks are measured by
the peak quantizer 9 which normalizes the signal samples within a time
frame. Forward error correction and dynamic bit allocation are applied to
the quantized samples by forward error correction generator 10 and by the
dynamic bit allocation technique or algorithm normally practiced in a
microprocessor as shown by the dynamic bit allocation section 11. The
output of the filter bank 5 is then companded or normalized in level by
compander 8 and quantized to the number of bits allocated by 11.
The output from the compander is typically a signal stream of approximately
13,000 bps and the forward error correction generator 10 generates an
output stream of approximately 3000 bps including the peak quantizer data
which presents a total data stream to the serializer 13 for transmission
over the digital channel 14 of approximately 16,000 bps. This serial
signal stream includes the actual companded signal samples plus a side
channel of information that indicates the bit allocations provided for
each frequency sub-band plus the forward error correction code.
Referring to FIG. 1 as the prior art, the 0 to 4 kilohertz input spectrum
on line 1 is typically sampled at 8000 samples per second after it emerges
from the A to D converter 2. This is shown by the sample clock 4
controlling the output on line 3 from the analog to digital converter 2.
Filtering and sub-sampling are conducted in the parallel filter bank 5
which decimates the incoming series of samples in the total spectrum into
sub-spectra, typically 8. In the example given, the 0 to 4 kilohertz input
spectrum is decimated into 8 sub-bands of 500 Hertz width each. The first
band is the 0 to 0.5 kilohertz band, the second is the 0.5 to 1 kilohertz
band, etc. Each of the sub-bands individual time waveforms are represented
by a 1000 sample per second bit stream at the output of the filter bank 5
as controlled by the clock 6. Numerous other bandwidths are sometimes used
and 16 bands of 250 Hertz width each are not unusual. Occasionally non
equal sub-band widths are employed.
The eight individual sub-band time waveforms are normally processed in time
block lengths ranging from 4 to 32 milliseconds in a signal processor
typically embodied as a microprocessor. The illustration in FIG. 1 assumes
a 16 millisecond sample block time length. The peak quantizer 9 in FIG. 1
finds the peak magnitude of the signal in each sub-band within a given
time block or series of samples. The individual sub-band frequency peaks
are logarithmically quantized, typically to a degree of 2 to 4 dB of
resolution.
The information is then passed to a dynamic bit allocation means and to a
forward error correction coder which adds error protection. The result is
then passed to serializer 13 which multiplexes the actual bits from the
transmission stream of samples coming from the parallel filter bank 5. The
bit allocation has been assigned at a reduced level by the bit allocation
technique practiced in box 11. The reduction in bits is to the level
occurring in the compander 8 which also multiplexes in the side channel
information which informs the receiver of the specific bit allocation
employed during this sample block of 16 milliseconds.
In FIG. 1, the dynamic bit allocation function 11 assigns available
bandwidth bits for a given block of 16 milliseconds of time to individual
frequency sub-bands normally at the rate of 1 bit for every 6 dB of peak
signal.
A given frequency sub-band with twice the peak value of a second sub-band
would get one more bit than the second. A band with four times the peak
energy would get two more bits than the other sub-bands and so forth. In
practice, this ideal assignment cannot be achieved since a fixed number of
available bandwidth bits cannot be subdivided precisely in this manner
among all the available sub-bands. The actual process performs an initial
bit assignment which includes possibly some very large numbers including
negative numbers and fractional numbers. These are then rounded to
integers and limited to a minimum of 0 and a maximum of, perhaps, 5 bits.
This usually results in the wrong total of bits required for assignment so
that an iterative redistribution of bits is required. All of this is a
highly time and hardware consumptive process which provides less than
ideal accuracy.
The sample compander and quantizer 8 uses the quantized peak energy
information to compand or normalize the time waveform in each band. It
then quantizes each sample in each frequency sub-band with the number of
bits that are assigned by the bit allocation technique for that frequency
sub-band. All of the information used for companding and for the bit
allocation is made available to the receiver or demodulator at the far end
of the system so that it can reconstruct the original time waveforms and
pass them through reconstructive digital to analog filters to approximate
the original 0 to 4 kilohertz input signal. The receiver end is not shown
in FIG. 1 but may be seen clearly in the IEEE International Conference on
Acoustic Speech and Signal Processing, Vol. 1 cited above.
In this process, some degradation in speech quality will take place since
the available bits, i.e., the bandwidth assigned for this coder on the
transmission system may not be sufficient to precisely reconstruct the
input signal in its original form.
The problems associated with this type of system are primarily those of bit
allocation. The bit allocation technique described above tries to
approximate the required functions and to achieve an optimum signal to
noise ratio for a given allowed fixed number of bits or bandwidth provided
to the coder. The first problem is that signal to noise ratio will be
highly affected in each speech spectrum due to the fixed number of bits
per second assigned. Flat spectra will have very few bits assigned in all
of their sub-bands while sparse spectra will have many bits assigned to a
few of the higher energy sub-bands and will thus yield high signal to
noise ratios as compared with the low signal to noise ratio in the flat
spectra. Secondly, it has been observed that humans do not hear noise in a
signal proportional to the signal to noise ratio. Additionally, not all
humans hear the same and do not hear according to any known equations or
mathematical models. The output quality of a speech coder is evaluated by
experts on its subjective equivalent signal to noise ratio based on their
appreciation of the perceived quality of speech as it is reconstructed.
Usually their estimate of the overall signal to noise ratio is
dramatically different from the actual quantitative signal to noise ratio
that exists.
Some partial solutions have been offered to these aforementioned problems.
Allocating bits at a rate of less than 1 bit per 6 dB of input peak signal
has helped. Non-linear quantizers have also helped. Variable bit rate
assignment techniques can help if one can determine how to vary the bit
rate assignment. All of these attempts amount to a guess at how humans
actually perceive the quality of sound through hearing. All of these
techniques use some formula that is convenient or easily implemented as
opposed to whatever is truly needed.
OBJECTS OF THE INVENTION
In light of the foregoing known problems with sub-band speech coders, it is
an object of the present invention to provide an improved method of
reducing the total quantized band peak information into a finite and
containable number of states that sufficiently represent both the level
and spectrum of each time block of signals and to provide means for
assigning the bit allocation to the sub-bands consistent with a desired
objective overall voice quality output.
Yet another object of the present invention is to provide improved means
for bit assignment utilizing the signal level and signal energy level
distribution spectrum for each time block to access a table of bit
assignments necessary for a given quality of speech output.
Yet another object of the present invention is to permit multiple bit
assignment tables to be used to allow bit rate assignments to be traded
for overall perceived speech quality output either to accommodate a
decrease in the bit rate when more users are present or to provide higher
quality service for one or more users than that provided to others for a
given bit rate.
SUMMARY OF THE INVENTION
The solution to the foregoing problems which meets the mentioned objects of
this invention may be set forth in its most general form as follows. The
total quantized sub-band peak information is reduced into a finite and
containable number of states that sufficiently represent both the absolute
peak energy level and the spectral energy level distribution within the
sub-bands during each time block of signals to be processed. The energy
level and the spectral distribution represent a given permutation of
signal states that can be used as addresses for table look-up. The tables
that are accessed contain the required bit allocations for any desired
subjective speech quality output. Table values are assigned based on
subjective experimentation and multiple tables can be provided for
different qualities of performance versus average bit rate throughput.
Separate tables could also be provided for signals such as those for
modems which require only an objective level of quality.
Processing slot time was chosen at 4 milliseconds with an 8 kilohertz
sampling rate so that 32 samples will be produced by the split band filter
during each 4 milliseconds. Four signal samples will be thus available for
each sub-band and 8 sub-bands are assumed. The 4 millisecond time block
was chosen to reduce the block processing delay and to minimize a round
trip echo path for live telephone conversations. Longer blocks reduce the
total amount of side channel bit rate information required but cause a
greater echo delay which is of more concern.
A band peak for a block is determined by computing the magnitudes of the
four samples in each sub-band and selecting the largest band peak
occurring during the block in each sub-band. An overall block peak is also
determined by finding the largest of the band peaks in the group of
samples over the 4 millisecond block. The block peak is quantized to fall
within one of thirty-two levels, each 2 dB apart, i.e., the block peak
found is compared to a 32 level block peak energy scale and the given peak
measurement for the block is identified at one of the levels. The peak
measurement is then encoded as a 5 bit binary number which is the number
of that level out of the 32 levels possible which has been found for the
block peak. This 5 bit binary number is to be included in the side channel
information for transmission to the demodulator to identify the overall
scale of the maximum peak in this block of data.
The overall block peak or absolute maximum energy level in the block is
then used to retrieve a multiplier or companding value to be used as a
scale factor to compress all of the samples from the various frequency
sub-bands to a normalized range. The companded samples are then quantized
in a linear 6 bit quantizer and temporarily saved as quantized samples.
The values of the samples will be 64 integers ranging from -32 to +31 and
including 0. Six bits of information is more than needed to represent many
of the samples and a reduction is made later.
The individual sub-band peak energies are also companded by the block peak
scale, but they are then linearly quantized with twice the resolution as
the individual samples. Band peaks are only positive numbers, i.e.,
magnitudes, so there will be only 64 levels possible. These are designated
levels 0 through 63 with 63 being the largest level.
Considering the nature of binary numbers, it is obvious that the bands
having peaks greater than 31 will require 6 bits of data to represent or
describe the level at which that sample fell on the scale from 0 to 63.
Those frequency sub-bands having peaks ranging on the level 16 to 31,
however, will have their two high order bits the same and the sub-band
frequency peaks lying in the range of 8 to 15 will have the three high
order bits equal and so forth. Thus, a table can be made to be addressed
quickly based on the individual band peak energy measured that will
indicate how many of the original 6 bit of quantized sample information
must be preserved to retain all of the original information contained in
the sample. Band peak levels lying on levels 32 through 63 require 6 bits;
band peak energy levels between 16 and 31 require 5 bits of the original 6
to be preserved, levels 8 through 15 require 4; levels 4 through 7 require
3; levels 2 and 3 require 2; level 1 requires only 1 bit and level 0
requires none.
The designation of the number of bits required can be encoded utilizing 3
bits since 3 bits can represent the numbers 0 through 7 which is more than
sufficient to encode the required bit designations 0 through 6. This
information is for the side channel of information to be passed to the
decoder at the receiver. With eight frequency sub-bands, allowing 3 bits
for each frequency sub-band to inform the receiver of how many bits in the
sample stream must be allocated to encode the individual band peak levels,
given 8 frequency sub-bands with 3 bits each, there will be 24 bits
required in the side channel for telling the decoder or demodulator how
many bits are required for each sample in each band. This is only an
indication of where the high order bit of each sample will fall.
Further bit rate reduction which will detract from the quality but not to a
detrimental degree will result based upon bit reductions applied uniformly
to the sub-bands based upon the findings of the energy spectrum and
absolute overall energy level in the time block sample. The bit rate
reductions to be applied are contained in a quality table which will be
described in greater detail later.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 illustrates a schematic prior art representation of a sub-band
encoder that can utilize the bit allocation technique of the present
invention.
FIG. 2 illustrates a truncated table showing schematically how the spectrum
of energy levels in the individual sub-bands and the absolute energy level
in the overall block of samples can be utilized as table addresses to
address a table containing bit allocation states that define how many bits
to be dropped from each signal sample.
DETAILED SPECIFICATION
As noted in the above summary, the coder as described therein can give
nearly transparent speech quality for telephone service. The bit rate
required will be excessively high however, reaching over 50,000 bps during
idle periods. To reduce this required bit rate, with very little
subjective speech quality degradation, it is possible to totally drop the
eight sub-band. However, it has been retained in the present
implementation since it may be required to meet some international
requirements. However, further bit rate reductions can be made based upon
the overall spectrum and energy level of the individual sample blocks as
will be described.
As described so far, the speech coder has within a block eight sub-bands
according to frequency, each with something between 0 and 6 bits of data
allocated to it to describe the block peak level lying in a range of 0 to
31 that can be encoded in 5 binary bits. Much of the time during normal
speech conversation, there are many more bits present in the samples than
are really necessary to preserve even the very high quality present. In
most cases, some of the bits could be dropped, particularly the low order
bits which represent only small analog output variations when the digital
signal is reconstructed. The low order bits or at least some number of low
order bits could be dropped while still preserving the overall general
high level of quality of speech reconstruction. The conditions in which
these bits could be dropped may be distinguished by the side channel
information, i.e., from the block peak maximum and from the individual
sub-band peak spectral distribution that shows the distribution of energy
across the spectrum in the sample.
A table can be constructed with a unique address for each of the side
channel state combinations with entries in the table arbitrarily
established on a subjective basis set by the quality of acceptable output
speech reconstruction desired. The table entries contain a number of bits
to be dropped for each state condition. The full number of addresses would
exceed twenty-six million, however, even if the eighth sub-band is
dropped. With seven sub-bands alone and with each sub-band having up to
six bits necessary to identify the energy level, there are 832,542
combinations in sub-band energy spectrum distributions possible with seven
sub-bands of frequency. With an overall energy level spectrum in the block
of 32 possible energy levels the total table size would be 26,353,344
entries. Such a table of entries would be cumbersome to say the least and
a number of steps are taken to reduce the number of entries down to 148
possible combinations for the present implementation.
The table can be truncated by combining or grouping some of the frequency
sub-bands together. Eight sub-bands or seven, are more than a sufficient
number to provide a good indication of the overall characteristics of
speech. Therefore, to reduce the size of the table and the consequent
addressing problem, the eight sub-bands are grouped into three sub-band
groups following somewhat the well-known articulation index. Band group 0
will contain the frequency sub-bands 0 and 1. Band group 2 will contain
frequency sub-bands 2 and 3 and band group 3 will contain the remaining
sub-bands 4, 5, 6 and 7. Each band group will be assigned the value of the
largest bit allocation in its included bands.
Further reduction in the size of the table can be achieved by reducing the
number of levels or bits to be assigned in each band group. The initial
levels were 0 through 6 bits. These may be represented differently by
grouping where any of the original levels 0, 1 and 2 can be called new
level 0. The original levels 3 and 4 can be represented as 1 and original
level 5 can be represented as 2 with original level 6 being represented by
a 3. No bit allocations are being redefined at this time, only grouping of
the information into more coarse segments is being carried out to reduce
the number of states to be addressed in the table.
These reductions have created a situation in which only three band groups
exist and each has only four possible levels. Assuming all combinations
were possible, there would still be 64 states or permutations of this
information. Only 37 states would have to be used if it is assumed that at
least one band group will have 6 bits allocated with its assigned level of
3. At very low levels of energy it is possible that no sub-band would have
all 6 bits allocated since the level of signal would fall below the
companding level. In this case, accuracy would not be very critical so it
is assumed that the level in band group 2 is at a level of 3 to thus
maintain the use of only 37 different permutations out of the original 64
mentioned above. These 37 states can be utilized as table addresses based
on spectra since they represent, in a condensed form, an indication of the
peak level distribution among the original seven or eight frequency
sub-bands, i.e., the spectrum of energy in the original block of samples.
A quality table constructed utilizing these 37 states as spectrum addresses
is of limited size and is of importance since human hearing is based on
the spectral distribution of energy in the signal heard. However, there
still remain 32 levels possible for the block peak. If all of these levels
were preserved and we would have 37 times 32 addresses in each quality
table for a total of 1184 addresses. The level is significant in the human
subjective speech perception quality only to the extent that human hearing
has a limited range. This means that perception of distortion is less at
lower sound levels. In the present specific implementation, the 32 levels
can be subdivided into only four general ranges. These can be subjectively
selected through a level table. Dropping the least significant bit of the
5 bit block peak reduced its range to only 16 levels with 4 dB separation.
The remaining 4 bit binary number can be used as an address to a 16
address table containing only the numbers 0, 37, 74 and 111.
To obtain the complete address, a level address is found using the block
peak that was measured, i.e., the 5 bit block peak with its least
significant bit dropped is a 4 bit number that can be correlated to one of
the four ranges noted above. This portion of the address plus the spectrum
address generated as pointed out earlier can be the quality table address
entry pointer. The quality table itself would then have 37.times.4 total
stored values, or 148 address locations. The quality table will contain a
value representing the number of bits to be dropped from each sample
actually coming from the parallel filter bank. The number of bits to be
dropped is ideally identified through preference testing with a group of
human subjects. Each address or selected groupings of addresses can have
the drop bit number varied while human subjects listen to phrases of
speech encoded utilizing the technique of this invention. Alternatively,
each address could hold 8 unique drop bit numbers, each corresponding to
each of the 8 sub-bands. By comparing the subjective opinions of the
listeners, each given address can have the drop bits maximized while
maintaining at least a desired level of quality. Multiple tables of 148
addresses can easily be generated each with a different quality level.
Four tables is a sufficient number to cover a reasonable range of near
transparency in which no distortion is perceptible to a minimum acceptable
quality that would require only a minimum bit rate.
In this fashion, the quality table supplies a solution to the complex
problem of how a human auditory nerve and brain interact to define a
listener's subjective evaluation of the coded speech performance. Previous
solutions have been based upon definitions that use some fixed equation to
optimize a few parameters including signal to noise ratio. Instead of
assuming a simplistic fixed equation, the quality table approach allows a
large group of listeners to define how bits should be removed from speech
blocks to optimize the subjective desires of speech quality in the group.
The quality table is thus a statistically defined speech coder and the
definition for resolution provided is based upon the subjective results of
a large body of listeners selecting the preference between two different
quality tables.
FIG. 2 illustrates schematically the table approach and shows how it can be
truncated from all possible combinations that result from 7 sub-bands in
the range 0 to 3500 hertz, the remainder from 3500 to 4000 being dropped
for this example, and 32 absolute block peak energy levels. Each table is
addressed by the 7 sub-band energy level numbers found in the sub-band
peak energy determining phase. The other entry is the overall absolute
energy level which falls within one of 32 levels previously noted.
The entire table may be condensed to a table of 148 entries following the
observance of the following rules. Step 1: the maximum of the sub-band
numbers 0 and 1 is called band group 0. The maximum value of the sub-band
peak numbers in sub-bands 2 and 3 is called band group 1. The maximum of
the band numbers 4, 5 and 6 is called band group 2. The coarse groupings
noted above are implemented in this table given a raw spectral
distribution as follows for the energy peak levels in the sub-bands.
Energy peaks might be represented as 4, 6, 2, 1, 3, 1 and 0 for the energy
sub-bands 0 through 6 and the absolute block peak energy level might be
decimal 4 or 00100 in binary. Assuming the rule of grouping the sub-bands
together, four arbitrary states A, B, C or D exist for each band group and
are assigned as follows. State A if band group 0 is 0, 1 or 2; state B if
band group 0 is 3 or 4; state C if the level of band group 0 is 5 and
state D if the level of band group 0 is 6. The same is true for the other
band groups 1 and 2. It will be recalled that band group 0 comprises the
measurements from the energy sub-band peaks 0 and 1, that band group 1
comprises the energy peaks from frequency sub-bands 2 and 3 and band group
2 comprises the energy peaks from frequency sub-bands 4, 5 and 6.
The absolute energy level also can occupy several different states which
were reduced to 4 states instead of 32 by a grouping as outlined earlier.
Values of absolute energy level from 0 through 7 are assigned to state A.
State B comprises the level 8 through 15; state C comprises the absolute
energy levels from 16 through 23 and state D comprises the energy levels
24 through 31.
Returning to the example begun above where a raw spectral distribution of
4, 6, 2, 1, 3, 1, 0 with an absolute energy level peak in the block as 4
being found, then the unique energy state defined in table 2 will be: Band
group 0, which is the maximum of 4 and 6, will equal 6 and corresponds to
state D. Band group 1, having a maximum of 2 and 1 will be 2 which
corresponds to state A. Band group 2 having the maximum of 3, 1 and 0 will
be 3 which corresponds to state B. The absolute energy level which is 4
corresponds to state A. Thus, the final location within the table is that
defined by the coordinates DABA. It will be noted that numerous spectral
band number permutations and numerous energy level permutations are thus
grouped together within a single region in the table and any specific
entry based on a permutation of the seven energy sub-band peak level
measurements and the absolute energy level measurement in the overall
block will fall within a given range and be given the same value within
the table. The content at the addressed portions of the table will be an
arbitrary number of bits, say 0, 1, 2 or 3 which are to be dropped from
the samples to be transmitted.
For the highest quality speech transmission, the high quality table indices
would be loaded with many 0 content locations indicating that no bits are
to be dropped from the samples assuming that the original signal samples
are the usual 16 bit 2's complement samples. For each of the four signal
samples in each of the frequency sub-bands, it is possible to greatly
reduce the overall transmission channel loading by dropping the least
significant bits and indicating to the receiver how many of the least
significant bits have been dropped from each sample during a given block
transmission period. Two or three bits being dropped from each sample may
create little or no noticeable distortion in the final received and
reconstructed signal but deleting even one of the high order bits will
produce serious distortion. The level of distortion that is acceptable is
a subjective criteria based on the acceptability of the overall perceived
signal to the human subject. A typical quality table of minimum acceptable
resolution would require deletion of 0 to 6 bits from each sample while
the highest quality table possible would dictate 0 bit deletion from each
sample.
It may be seen that quality tables may be established for any range of
quality possible from the highest to the lowest acceptable quality level.
It is equally clear that operation on a specific voice line to a specific
user might be dictated as being accorded a high level of quality for
priority communications but only a very low objective level of quality
would be provided for machines such as a modem over the same line. The
only difference in construction would be that different quality
performance tables would be utilized depending on the nature of the input.
A user wishing to acquire service over such a communication system might
opt for high quality, high signal to noise transmission and be assigned a
performance quality table with little or no bit deletion at a relatively
higher price since proportionately more of the total channel bandwidth
will be utilized in supplying high quality service and the consequent
number of bits required to generate the high quality will occupy more of
the channel space. Similarly, the user might request the least expensive
or least proportionate amount of channel bandwidth for use by a modem.
Indeed, the range of possible services and qualities thereof is virtually
limitless if one assumes the full possible range that these quality tables
can assume.
A more usual scenario would be that the system such as shown in FIG. 1
would implement the highest quality service for all users consistent with
the total demand placed on the system by the number of users present given
the channel bandwidth available to the system. Should either the channel
bandwidth become restricted or reduced or the number of users increased,
resort to lower quality transmission tables, i.e., assignment of fewer
bits to each user can be easily accommodated by switching to a different
speech quality table for the next and any ensuing sample blocks until
either the user demand decreases or channel capacity is increased.
Examples of quality tables are empirically determined as noted above. Two
examples follow, one for a moderate degree of quality and one for a low
quality. Spectrum addresses for the table are generated as stated
previously to generate one of 37 possible states. A simple rule for
address generation is: If band group 0 has a band group level value of
less than 3 and band group 1 has a level value less than 3 then the
spectrum address will be three times the band group 0 level plus the band
group 1 level which will generate spectrum addresses 0 through 8. If band
group 0 has a level less than 3 and band group 1 is a level of 3, then the
spectrum address will be four times the band group 0 level plus the band
group 2 level plus 9 which will generate spectrum addresses from 9 to 20.
And if the band group 0 level equals 3, then the spectrum address will be
four times band group 1 level plus the band group 2 level plus 21 which
will generate the spectrum addresses from 21 to 36.
The other entry to the table is from among the 32 levels of the block peak
that were possible. The 32 levels were condensed to 4 ranges by dropping
the least significant bit of the 5 bit block peak level which reduces its
range to only 16 levels of 4 dB separation each. The remaining 4 bit
number is used to address a 16 address table containing only the numbers
0, 37, 74 and 111 as noted earlier. The actual extent between the dividing
points in the table between the values 0, 37, 74 and 111 is determined
empirically by letting users perceive the result and select the most
acceptable perceived quality. However, it may be observed that with high
energy levels, the overall perceived quality is highly subject to signal
to noise ratio distortion and that at very low levels the signal to noise
ratio is not very important in a perceived quality of speech. Thus,
dividing the four possible ranges equally among the possibilities would
assign a higher degree of quality to very low level signals than is
necessary. Consequently, in the example given, only two of the sixteen
possible addresses are assigned to the lowest energy states, three to the
next lowest, five to the next and six to the last or highest energy level
range. Thus, in the specific implementation proposed herein, instead of
utilizing the full 32 levels that are possible for the absolute energy
level peak, the 32 levels are divided into only 4 ranges and the ranges in
which a given absolute energy level measurement falls will dictate a range
of table values to be accessed.
As a review of the overall coding process, the steps are as follows. First,
the incoming signal samples are divided into blocks. In our example, all
the samples occurring within a 4 millisecond period are a block. The peak
within the block, i.e., the peak magnitude of all 32 signal samples in the
8 sub-bands with 4 signal samples per sub-band is found. Thirty-two
logarithmic sized segments are established for the range of the block
peak. The segment in which the given block peak is found is encoded as a 5
bit binary number representing the segment among the 32 in which it was
found. Next, a signal sample range is found lying within the upper bound
of the block peak segment that has been identified for the given block
peak. Sixty-four linear sub-segments are defined for the range in which
the block peak lies and then a specific sub-segment in which a given
signal sample lies is defined by a 6 bit 2's complement binary number to
identify, within the range of the peak, which l | | |
