Experimental Speech System, Part One published by Practical Electronics November 1983
Posted 27 october 2002
SPEECH SYSTEM
OVER recent years some extremely sophisticated techniques have been developed
for the digital encoding of speech, and for its reconstitution. The purpose of
this article is not to discuss these techniques (see references),but to present
a very simple method, and its implementation using a 6502-based microcomputer
with minimal extra hardware, whereby speech may be digitised, and subsequently
regenerated, with an adequate degree of intelligibility.
This method is relatively economic in its use of memory for storage of the speech
data, and has the outstanding advantage that the encoding process is simple. One
may readily construct any chosen vocabulary, whether composed of complete words,
phrases, or phonemes. This is in distinct contrast to the most commonly used method
of speech digitisation, involving linear predictive coding to reduce redundancy,
and utilising a digital model (a lattice filter) of the human vocal tract for the
reconstitution of speech. Although highly complex, this hardware is now cheap and
readily available. However, it is not at all easy for the user to construct his
own vocabulary, as the encoding process is so complicated.
The method presented here is not new in concept and is based upon the well-known
observation that the intelligibility of infinitely clipped, or zero-crossed speech,
is remarkably high. This is due to the particular spectral characteristics of speech,
especially its quasi-periodic nature. Anyway, given this observation, it is but a
simple step to see that one may encode zero-crossed speech waveforms by storing
digitally the time intervals between successive zero-crossings.
The zero-crossed speech waveform may be faithfully reconstructed by an inverse
process of generating switching intervals corresponding to the magnitudes of the
digital speech data. Note also that speech may be speeded up or slowed down on
replay (but the pitch also changes accordingly). In practice it was not found
possible to speed up the speech appreciably on replay, because of the limited
speed of the microprocessor.
Initial experiments had the simple aim of digitising a few seconds of continuous
speech (including any periods of silence). Once this had been achieved, and the
reconstituted speech shown to be almost identical to the original zero- crossed
speech waveform, the programs and hardware were developed further to enable isolated
words or word segments to be digitised and stored between known limits in memory.
Knowing where in memory the speech data comprising a word startsand ends, it becomes
a simple matter to replay any particular chosen word under program control.
FIRST STEP:
Acquisition and replay of a few seconds of continuous speech, including silences.
A little consideration suggested that sufficient speed could be obtained only by
using the 6502 interrupt system; it would take rather too long to repeatedly examine
a digital input line, testing for a change of state. The acquisition hardware operates
as follows: First a source of audio signal with a peak amplitude of a few volts
rather than millivolts,is required.
This can be obtained from a microphone followed by a suitable preamplifier (Fig. 1.1.),
but if such a pre-amp is not available, then the audio signal may be pre-recorded and
replayed from a cassette recorder at maximum volume, taking the audio output from the
earphone socket. In Fig. 1.2, IC1 is configured as a comparator receiving an input
which is filtered to give a bandpass response between about 300 and 3200Hz. The
comparator threshold may be set within a certain range either side of zero by means
of VR1. Ideally, the threshold should be set at zero, but in practice it is preferable
to set the threshold just above the noise level, either positively or negatively.
Clearly it is necessary to conduct these experiments in an environment relatively
free of acoustic noise.
The comparator provides a "zero-crossed" output or, expressed another way,
an infinitely clipped version of the audio waveform. The function of the pre-filtering
is to restrict the bandwidth to the minimum compatible with reasonable intelligibility.
This will economise on memory usage for the storage of the words, especially those
with a high sibilant content. IC1 is followed by an edge detector, consisting of a
simple delay and an exclusive-OR gate. This gives positive-going pulses of minimal
duration (the acquisition Count routine must not interrupt itself).
These pulses are inverted by IC3a 1/2 74LS20 and then supplied to the IRQ input of
the 6502 microprocessor. The reason for using a 4-input NAND gate rather than a simple
inverter is to enable the system to be expanded later on.
ACQUISITION SOFTWARE
Table 1.1 gives the assembly listing. Table 1.2 gives a hexdump of the acquisition program.
The storage area reserved for speech data is $1000 to $1FF, so a minimum of 8K of memory
must be available. 4K of memory is thus available for storage of speech data, which is
enough for a few seconds of speech.
INTA1 is the acquisition program. It consists of two parts, an Initialisation routine, and
a Count routine. Superboard II has a small area of free memory, located at addresses $0250
to $02FF, which is not used by BASIC, so it was decided to assemble the programs into this area.
Note that the machine used was fitted with WEMON; on machines without WEMON this area of free
memory in page 3 is slightly larger.
The first part of the acquisition program is concerned with initialisation. It ensures that
all arithmetic operations are conducted in twos complement binary rather than in BGD; then it
sets up the system IRQ vector to access the Count routine. The remaining instructions set up a
2-byte indirect pointer to access memory for the storage of zero-crossing intervals, starting
at $1000; then the counter (the X-register) used for timing the intervals is cleared, interrupts
are made possible, and the processor cycles in a jump-self wait loop until the first interrupt
pulse arrives. Operation of the second part, the Count routine, is best understood by reference
to the assembly listing. Essentially it stores the last count value, and counts again until
another interrupt pulse occurs, in which case the counting starts again, or until a count of $FF
is reached, in which case the same thing happens, except this time via a JMP rather than an
interrupt. Thus intervals longer than $FF count units will be truncated to that value.
In practice most intervals do not exceed $FF units, but the reason for incorporating
this feature is to enable silences also to be encoded, as a train of bytes of value $FF.
For this first experiment this was of interest, because some words do contain brief silences,
and also to enable continuous speech to be encoded. Later it is attempted to record single
words one at a time, both with and without this feature, to see how great a difference the
omission of brief intra-word silences makes. For the purposes of the present programs, however,
it is enough to note that the replay program is designed to skip over bytes having the value
$FF, i.e. it treats them as silence. At the other extreme there is a minimum interval between
consecutive zero-crossings that can be correctly resolved; this is the time taken for the
execution of instructions in lines 180 through 300. Again, the quality of the resuts
obtained suggests that intervals so brief are rare in band-limited speech.
REPLAY OF ZERO-CROSSED SPEECH
The hardware required for reconstitution of the zero- crossed speech is even simpler than
that employed for acquisition. Best results are obtained if the replay is filtered to remove
the high-frequency raggedness of the sound, and the low-frequency rumble component. This is easily
achieved using an audio amplifier (Fig. 1.3.) equipped with bass and treble controls.
Alternatively, a simple audio amplifier may be used in conjunctionwith the same pair of filters
used for the acquisition process. The replay program, Table 1.3, generates pulses on a particular
write line, in this case W2, whose address is decimal 61314.
The pulses occur at intervals corresponding to the original zero-crossing intervals, and are used
to toggle a flip-flop, Fig. 1.4, to generate the reconstituted infinitely clipped speech waveform.
Clearly at least one address-decode write line is needed.
See references for suitable circuitry. The functions of capacitors C1 and C2 is not immediately
obvious, but they were found necesary to ensure clean toggling operation of the flip-flop.
I speculate that there may have been double-pulsing or ringing on the write line, although I could
not detect it by eye on an oscilloscope. Thus these capacitors may not prove necessary, and
I suggest experimenting with their values.
NEXT MONTH:
The second part of this article will deal with programs and additional hardware for building
up a vocabulary of isolated words, and for replaying specific words under program control. It will
also consider some of the possible uses of these techniques for producing complex sound effects
rather than speech, and look at zero-crossed speech data graphically with a view to possibilities
for' speech recognition.
REFERENCES
1) Speech Synthesis. Practical Electronics Nov., Dec. 1980
2) Interfacing Compukit, by D. E. Graham, Practical Electronics Jan.-July 1981.
Experimental Speech System, Part Two published by Practical Electronics November 1983
SPEECH SYSTEM
TWO approaches have been tried for the systematic acquisition of a vocabulary and for selective
replay of a specified word. They have both been considered because they have their respective merits.
In the first approach the acquisition count routine halts during silences; this has the great
advantage of ignoring redundant silence either side of a word as spoken during the period that
the Acquisition Enable button is depressed. However, intra-word silences are elided. One solution
to this difficulty might have been to make acquisition conditional on the state of the output of an
envelope detector; however, this would still be unsatisfactory, because the first few milliseconds
of a word would be missed, owing to the response time of the envelope detector. Furthermore,
the acquisition procedure would be fooled, since it would be unable to distinguish between such
a brief hiatus and the true end of a word.
Yet other approaches might be tried: obviously the problem could be overcome by using a double-precision
counter, which would time for long enough to cover the intra-word silences; but it would be wasteful of
memory to store two bytes per intertransition interval, for much of the time the high-order byte would
be redundant; it would also (in the case of the 6502) increase the count loop time, so making for poorer
resolution of intervals.
The first approach was tried, and was quite successful, apart from the elision of intra-word silences.
In the second approach, for which programs are here presented, provision is made for the acquisition
of silences of a limited duration, whose maximum value may be specified before execution of the
acquisition program. The end address of the speech data is automatically recorded in a table when the
press button switch is released. This is achieved by means of a routine to which control is transferred
by a negative-going transition on the 6502's NMI line. The replay program regenerates any chosen word,
the appropriate speech data being accessed simply by a number, this being the position of the word in
the sequence that was originally stored. Note that 16K of RAM is needed, and that BASIC must be restricted
to decimal 5567 bytes. The additional hardware needed has already been given in the lower part of Fig. 1.2
last month, but revised, rationalised circuitry is given in Fig.2.1.
REVISED ACQUISITION CIRCUITRY
The acquisition circuitry has been modified to avoid the need for a dual power supply, and to
incorporate some nominal filtering to restrict the speech bandwidth. If you wish to experiment with
different bandwidths, then omit this filtering and use the variable low and high-cut filters
described earlier. A further refinement is the provision of
an l.e.d. driven by the comparator output; this is helpful in setting the comparator threshold
just above the noise level: VR1 should be set so that the l.e.d. just stays off at the ambient
'silence' level. (It is important that background noise be reasonably low within the operating
bandwidth). A delay interposed between the comparator and the I.e.d. stretches brief threshold
crossings enough to visibly light the l.e.d.
REPLAY CIRCUITRY
Quite acceptable speech output can be obtained, without the need for an audio amplifier, by
connecting a small
loudspeaker to the flip-flop as shown. Placing the loudspeaker cone downwards in a small plastic
bowl of suitable tapering diameter was found to provide a beneficial resonance; this is something
to experiment with.
SETTING UP OF HARDWARE
Setting up the acquisition circuitry involves only adjustment of VR1 as already described.
To verify that the D-type flip-flop can toggle correctly, fill up part or all of the word data
storage area with some arbitrary interval value (except 255, which is treated as silence), and execute
the routine for continuous replay, i.e. RPFF3 given earlier. This should yield a continuous tone of
steady pitch; if it sounds ragged or irregular, then try changing the values of the capacitors marked
with an asterisk.
CONSTRUCTION
The circuitry is not complicated or critical in layout, and because it is regarded as experimental
and open to improvement, a p.c.b. design has not been provided. It can readily be assembled on Veroboard
or fitted into space on an existing board.
USING THE PROGRAMS
Having set up the hardware, the programs may now be tried out.
Enter a value into SILMAX, i.e. page zero location 51 hex. This gives the maximum length of continuous
silence that can be acquired; a value of 10 hex seems about right. Now execute the acquisition program
from 029B hex. No speech data will be stored in memory until the Acquisition Enable button is pressed.
Thus the procedure is as follows: decide on the vocabulary sequence you wish to enter; execute the program
from the machine code monitor; press the button and hold down, starting to speak the word as soon as the
button is depressed, and releasing the button as soon as the word is finished; continue in this fashion
until the program returns you to the monitor, which will happen when the memory storage area for either
the word data or the end-of-word address table becomes filled up. Depending upon the bandwidth in use, up
to 20 average-length words can be stored. If you have more than 16K of memory, then you can modify the
programs to increase the vocabulary storage capacity.
Having entered the vocabulary sequence, verify that page zero location 57 hex contains the number of
words which you were able to enter; then have a look at the contents of the table of end addresses
of words; the two-byte values in low-high order should be in an ascending order.
The replay routine is conveniently tested from BASIC using the test programs given, one of which
replays a single specified word, the other replaying the entire vocabulary out in sequence. Clearly the
replay routine can be called from any other program of your choice.
It is actually possible to manage without a microphone" and preamplifier. because you can store
the vocabulary sequence on cassette tape, and then replay the tape recording into the acquisition circuitry
at full volume, pressing and releasing the button switch as described above; a particular vocabulary for use
with a particular program can be stored on tape along with the program; clearly the vocabulary data could also
be stored in its encoded form, in which case the acquisition procedure would not have to be repeated. If you
have a disc-based system, then the speech facility becomes much more useful.
