| Chapter 1 | Setting up the CURRAH MicroSpeech | ||||||||||||
| Chapter 2 | Simple words and phrases | ||||||||||||
| Chapter 3 | The Allophone Set | ||||||||||||
| Chapter 4 | Intonation and "phrases libraries" | ||||||||||||
| Chapter 5 | The speech buffer - using CLEAR | ||||||||||||
| Chapter 6 | Using machine code with the unit | ||||||||||||
| APPENDICES |
|
The CURRAH MicroSpeech is an allophone speech synthesiser, which means that it uses individual speech sounds strung together to make intelligible speech. This is different to some speech synthesiser available today which have a fixed vocabulary; unlike these systems the MicroSpeech has an unlimited vocabulary and can synthesise any word or sentence in the English language. The user merely has to become familiar with he speech sounds of English (which are different from letters) and the allophone symbols used to represent them.
To set up the unit, remove it from its packing tray if you have not already done so, UNPLUG THE SPECTRUM FROM ITS POWER SUPPLY, and plug the unit into the port on the back of the Spectrum, CURRAH logo uppermost. You will find that there are two short leads at the back of the MicroSpeech unit:
Remove the Spectrum's TV lead from the "TV" socket on the back of the Spectrum and plug in the short UHF lead instead. Then take the Spectrum's TV lead and plug it into the Phono socket on the back of the MicroSpeech. This now allows speech to be "mixed" with the Spectrum's video signal so that it comes from the TV speaker.
Now take the line lead and plug it into the "MIC" socket on the back of the Spectrum (you cannot do any harm if you plug it into the "EAR" socket by mistake). This lead allows the sound generated by the Spectrum's own internal loudspeaker to come from the TV speaker - simply remove this lead when you wish to use a cassette recorder with your Spectrum.
Plug in the power lead to the Spectrum. The Spectrum will initialise
and the message:
"Speech System (c) CURRAH 1983"
will appear at the top of the screen, as well as the normal Sinclair Research Ltd. copyright message. If this does not happen, unplug the power, check that you have connected the unit up correctly, and try again.
Turn up the volume on your television (about half-way will do) and press ENTER. The unit will say "enter". If it does not, there are two simple checks you can do:
Now try pressing any of the keys (remember to press ENTER first if you have just switched on). All BASIC keywords and printable characters are voiced, (with the exception of the "~" character) so when LET appears on the screen you will hear the word "let" - go on - try it! Note that all graphics characters are voiced as "graphic", whatever their shape.
All keyvoices "repeat" - if you try holding down ENTER for a few
repeats, the unit will say something like
"E...E...E...E...Enter".
The old keyvoice is "chopped off" and a new one started each time the repeat occurs. If you type very fast the same sort of effect is obtained.
Try all of the keyvoices now. Apart from being useful in themselves, the keyvoices serve to illustrate what can be achieved by allophone speech synthesis.
When you are bored with this, go on to the next chapter. If at any point you get stuck, remember that all you have to do is unplug the power, wait a moment, and start again - the unit initialises automatically each time.
You may keep the keyvoices on as long as you want, but you may find
them a nuisance after a while, so enter:
LET keys=0
And you will find that the keyvoices have vanished. Try entering:
LET keys=1
And they reappear again. What you have done is to use a "reserved
variable" to tell the unit what to do - the variable "keys" is
"reserved" for the unit's use. You can put this in a program line, too
- try entering this program:
10 LET keys=0
After running this, the keyvoices are disabled again.
Once you have enabled or disabled the keyvoices, they remain either on or off until you enter a "LET keys=" command, either in a program line or entered directly as a command (note that the keyvoice status is preserved even after a NEW). There is another reserved variable, s$, (standing for "speech string") which is used for generating user-defined speech.
Try entering the following:
LET s$="he(ll)(oo)"
You will hear the unit say "hello" (if you do not, type it in again, checking that you get all the brackets in).
Now try entering this program:
10 LET s$="he(ll)(oo)": PAUSE 1
Because the way in which the MicroSpeech works, you MUST put a PAUSE statement after each "LET s$=" statement in a program to make sure that the unit detects each s$ as it occurs (PAUSE 1 will do).
If you RUN this program, you will again hear the word "hello". Every time you create a new s$, the contents of the string are examined by the allophone interpreter and, if the syntax is correct, the allophone symbols are converted into speech code and placed in a "buffer" (a sort of queue) near the top of memory. The code in the buffer is then automatically outputted to the speech chip without any further intervention from your BASIC program.
You BASIC program can go away and do other tasks once the speech has been stored, so you can fill the buffer up and leave it to get on with outputting speech whilst you type in some data, for instance, or add more speech to the buffer.
User-defined speech has higher priority than key voices, so that if a sentence is being spoken and you press a key, the keys will not voice, even if they are enabled, until the buffer has been emptied.
The unit has a self-diagnostic facility which reports on the status of
the speech you tried output. RUN the above program again, and then
enter:
PRINT s$
The computer will report back with:
*e(ll)(oo)
The asterisk obliterating the first character tells you that the
allophone symbols contained in s$ were syntactically correct and that
they were converted into speech code and loaded into the buffer for
output. Now try changing the program to read:
10 LET s$="he(ll)!(oo)": PAUSE 1
No speech will be generated if you RUN this as the exclamation is a
"illegal" character. RUN the program, then type in:
PRINT s$
And you will get the following report:
?e(ll)!?oo)
The query (?) obliterating the first character tells you that there was an error in the string, and the second query shows you where the error occured. In the case of an error occuring on the last character of the string, the second query obliterates it.
If you BREAK into a program whilst the voice is speaking the voice ceases at once - what is not so obvious is that any "report" given by the Spectrum other than "OK" will also have this effect: e.g., STOP statement, Out of DATA, RETURN without GOSUB, etc. This is to ensure that if there has been an error in a program which may have left you with a buffer full of data you will not have to wait for the buffer to empty itself - remember a buffer only 200 bytes long can hold over 30 seconds of speech.
You have seen four examples of the allophone symbols used already, and you will probably be wanting to try constructing you own words and phrases. When you feel that you are familiar with the use of the reserved variables, go on to the next chapter, where the whole allophone set will be introduced.
As you may have noticed, there are basically two sorts of allophone symbols, those that occur as single characters and those that have several characters enclosed by brackets. It is necessary to enclose some in brackets because "aa" for instance, is a different sound to "(aa)".
Here is a complete list of all the allophones. Try them to hear what
they sound like by running this short program:
10 INPUT a$: LET s$=a$: PAUSE 1
20 IF s$(TO 1)="?" THEN PRINT "You have made a mistake"
30 IF a$="stop" THEN STOP
40 GOTO 10
Now whenever you come across an allophone you would like to hear, simply enter it when the "L" prompt is given and the unit will say it for you. In the following table, the allophone symbols are shown on the left, with examples of the sounds on the right:
| All single characters | All sound phonetic |
| a-z excepting x and q | e.g. like "a" in at, "b" in tab, etc. |
| (aa) or (ay) | "ay" in hay |
| (ee) | "ee" in see |
| (ii) | "i" in hive |
| (oo) or (eau) | "o" in stove |
| (bb) | "b" in bat |
| (dd) | "d" in do |
| (gg) | "g" in got |
| (ggg) | "g" in big |
| (hh) | "h" in hoe |
| (ll) | "l" in let |
| (nn) | "n" in no |
| (rr) | "r" in run |
| (tt) | "t" in to |
| (yy) | "y" in yeah |
| (ar) | "ar" in arm |
| (aer) | "air" in repair |
| (ch) | "ch" in church |
| (ck) | "ck" in clock |
| (ear) | "ear" in clear |
| (eh) | "ar" in wary |
| (er) | "er" in leader |
| (err) | "ur" in purr |
| (ng) | "ng" in tongue |
| (or) | "or" in sore |
| (ou) | "oo" in root |
| (ouu) | "oo" in food |
| (ow) | "ow" in now |
| (oy) | "oy" in boy |
| (sh) | "sh" in ship |
| (th) | "th" in thin |
| (dth) | "th" in then |
| (uh) | "oo" in took |
| (wh) | "wh" in whig |
| (zh) | "z" in azure |
| apostrophe | very short pause - e.g. in can't |
| space | pause between words |
| comma | pause between phrases |
| full stop | pause between sentences |
X and Q are missing from the set as the speech sound they make may be made from combinations of other allophones - this is the key to allophone synthesis. For instance, the word "queen" may be made by "kw(ee)n" and the word "box" by "boks".
Certain allophones run together well, for example:
LET s$="aaaaaaaa"
Comes out as a single "Aahhh" sound. Other allophones that exhibit this property are e, i, o, u, f, s, (eh), (th) and (uh).
To help you see how words are built up, here are some examples from the keyvoices:
| Keyvoice | Voiced by using: |
| AT | a(tt) |
| ' | apostruf(ee) |
| = | (ee) (ck)wulz |
| NEW | ny(ouu) |
| SIN | s(ii)n |
| CODE | c(oo)d |
| ACS | (ar)c'coz |
| STRING | stri(ng) |
| CIRCLE | s(er)cul |
| LLIST | e(ll)'list |
| SAVE | s(ay)'v |
| 5 | f(ii)v |
| H | (ay) (ch) |
| LOG | lo(ggg) |
| THEN | (dth)en |
| MERGE | m(err)dj |
You can see that a little care is needed to think in terms of how words are spoken rather than how they are written, but with a little practice you will be able to build up words very fast with few mistakes.
Try making up some words of your own now, and making short phrases using the various pauses to divide the words up. You should also try LISTing the BASIC demonstration program supplied with the unit to see how the various sentences are generate on that.
Intonation is necessary in allophone speech to add characters to words or parts of words. Up to now you will have been entering speech strings in lower case letters. If you use UPPER case letters of any allophone, the intonation goes UP. Try this:
LET s$="aaAAaaAAaaAAaaAAaa"
And you will hear a sound rather like someone yodelling. Note that the pitch of the allophones does not go up or down suddenly - it is "ramped" so that the transition is smooth. Try This:
10 LET s$="sp(EE)k n(oo) (EE)vil": PAUSE 1
You will be able to hear the intonation. Note that the intonation will be most noticeable on vowels and "voiced" allophones, but try experimenting with it to see what sort of effects you can obtain.
The keyvoices have relatively little intonation added so that they sound "neutral" for the purpose of typing in programs.
In a bracketed allophone, it does not matter if you mix upper and lower case inside the bracket, as long as you realise that it is the LAST character inside the bracket which determines the intonation of the allophone. For example, "(oUu)" is not intoned, whilst "(ouU)" is intoned up.
When you are using the MicroSpeech unit to make sentences, think how
much easier it would be if you could build up sentences from a
"library" of words and phrases. Suppose you wanted to make the unit
say "I'm sorry, but your answer was incorrect". Rather than put this
in one long string, you might want to construct the sentence from a
"library" of stock phrases, like this:
10 LET a$="(II)'m sor(ee)"
20 LET b$=",but"
30 LET c$="y(OR) ans(er) woz"
40 LET d$="incurrect"
50 LET s$=a$: PAUSE 1: LET s$=b$: PAUSE 1: LET s$=c$: PAUSE 1:
LET s$=d$: PAUSE 1
Whilst this will work very well, there is another solution. You could
set up the four strings as before but join them together
("concatenate") them into s$ by changing line 50 to read:
50 LET s$=a$+b$+c$+d$: PAUSE 1
Whilst this is more compact than the first method, it is still rather
wasteful of strings. If you have to build up sentences from a library
of more than ten phrases, then you will probably find that setting up
a string array and concatenating the elements into s$ will be a better
solution. You will need to "slice" some elements before concatenation
in order to suppress the trailing spaces:
10 DIM a$(4,18)
20 LET a$(1)="(II)'m sor(ee)"
30 LET a$(2)=",but"
40 LET a$(3)="y(OR) ans(er) woz"
50 LET a$(4)="incurrect"
60 LET s$=a$(1)( TO 14)+a$(2)( TO 4)+a$(3)+a$(4)( TO 10)
70 PAUSE 1
It was explained earlier that speech data is stored near the top of memory until it is ready to be outputted.
The speech buffer is a "First In - First Out" buffer (a FIFO buffer) which is initialised when you turn on the Spectrum with the MicroSpeech attached. In order to do this, the unit moves the "top of BASIC RAM" pointer (RAMTOP) down by 256 bytes (so that on a 48K machine it will point to 65111) and the speech buffer is then established between the new RAMTOP and the user-defined graphics area. The buffer then fills from top (i.e. highest memory location) downwards as more speech is added to it. The top 6 bytes of the buffer (the "header") contain information on buffer and system status but the ones below it are free for speech data. You may make the buffer as large or as small as you want by use of the clear command.
For instance, on a 48K machine, CLEAR 65000 makes RAMTOP point to 65000, and the memory space free for the speech buffer will be increased by 111 bytes.
To safeguard against the buffer header being corrupted by a CLEAR command, the MicroSpeech unit monitors RAMTOP, and if a CLEAR is attempted into an illegal area (see below), the unit goes through its initialisation sequence again.
You BASIC program is not lost when this happens (unless it was so large that the new RAMTOP established by the unit truncated it), but you will find a new copyright message issued and RAMTOP once again established 256 bytes below the original RAMTOP.
If you try to add speech to the buffer and there is insufficient room for it to be fitted in, then the unit will ignore the speech string s$ until there is sufficient room in the buffer. If you PRINT s$ under these conditions you will find it returned uncorrupted, i.e. without an asterisk or a query, to show that it has not been accepted into the buffer. If you find this occuring in a program, simply use CLEAR as described above to make the buffer a little larger, or alternatively use a PAUSE to allow the buffer to empty a little before attempting to add another string to it.
Since the keyvoices are outputted via a speech buffer, you can observe a similar effect if you enable the keyvoices when the buffer is very small. For instance, if you make the buffer only four bytes long, then any keyvoice longer than four allophones will not be voiced.
Summarising:
On a 16K machine, do not CLEAR greater than 32343
On a 48K machine, do not CLEAR greater than 65360
This section is intended to set out guidelines for writing machine code programs for the unit. It must be stressed that the MicroSpeech was primarily designed to be used with BASIC for wider popular appeal, and as the operating software is necessarily complex only the experienced programmer should attempt writing machine code for the unit.
In the following examples, the 48K version memory address will be used - simply subtract 32768 to convert to the 16K machine. Decimal is used for all numbers unless otherwise stated.
At the top of the speech buffer is the six-byte "header" - information which controls the action of the buffer and the speech system. This comprises:
| ADDRESS (48K) | FUNCTION |
| 65367 | Flag byte |
| 65366 | Spare byte |
| 65365 | Hi byte of buffer pointer |
| 65364 | Lo byte of buffer pointer |
| 65363 | Spare |
| 65362 | Spare |
| 65361 | Next allophone to be voiced |
The Buffer Pointer keeps a pointer to ONE BELOW the last item in the buffer which was voiced. It is incremented by one every time an allophone byte is outputted to the speech chip from the top of the buffer and the entire contents of the buffer is moved up by one ready for the next byte to leave. If this pointer points to 65361, as it does at startup, then the buffer is deemed empty and no allophone is issued.
The Spare bytes are initialised to zero (00h) at startup and serve as safe locations for you to store vital data in - you may want to experiment with a dual buffer, for instance, and switch between the two by swapping pointers.
The Flag Byte contains various flags for system use.
The function of each bit is as follows:
| BIT | FUNCTION |
| 7 | Copyright message flag - reset after issuing message |
| 6 | Spare |
| 5 | Spare |
| 4 | Used in keyboard scan - unlikely to be useful |
| 3 | Spare |
| 2 | Used in keyboard scan - unlikely to be useful |
| 1 | Key voice enabled (set at first) |
| 0 | Spare |
None of the flag bits will cause a crash if you alter them although the copyright message will appear if you set bit 7.
The important thing NOT to corrupt is the buffer pointer - if you make this point below the current RAMTOP you will crash the system.
If you would like to input speech data directly into the buffer rather than go via the BASIC variable s$ you must first fill the buffer with the allophones you want.
Each allophone in the buffer is represented by a byte; six bits of each byte are devoted to the allophone code, one bit is for intonation, and bit 7 is always zero.
| BIT: | 7 | 6 | 5 4 3 2 1 0 |
| FUNCTION: | zero | intonation | allophone |
| (1 for up) |
The decimal and hexadecimal codes for the allophones are given in appendix I. As an example of machine code programming, let us suppose that you wanted to make the unit say "hello". First of all, you must work out the codes for "he(ll)(oo)" - they are 27, 7, 62 and 53.
Let us also suppose you wanted the "e" to be intoned up. Simply add 64 to the basic code (7) to give 71. Now that you have your data, you have to load it into the buffer in the right order (top down) and then update the buffer pointer to point to ONE BELOW the last allophone you wish to be voiced. Here is a suggested program which would do this, written as an assembly language subroutine:
Start: ld hl, Data
ld de, 65361
ld b, 4 ; load count with 4
Loop: ld a, (hl)
ld (de), a ; load allophone in buffer
inc hl
dec de
djnz Loop ; loop to load four bytes
ex de, hl ; point hl to 1 below the last byte in buffer
ld (65364), hl ; update pointer
ret
Data: Defb 27, 71, 62, 53 ; the four bytes data
| ALLOPHONE | DECIMAL | HEX | ALLOPHONE | DECIMAL | HEX | |
| a | 24 | 18 | (aa) or (ay) | 20 | 14 | |
| b | 28 | 1c | (ee) | 19 | 13 | |
| c | 8 | 08 | (ii) | 6 | 06 | |
| d | 21 | 15 | (oo) or (eau) | 53 | 35 | |
| e | 7 | 07 | (bb) | 63 | 3f | |
| f | 40 | 28 | (dd) | 33 | 21 | |
| g | 36 | 24 | (gg) | 61 | 3d | |
| h | 27 | 1b | (ggg) | 34 | 22 | |
| i | 12 | 0c | (hh) | 57 | 39 | |
| j | 10 | 0a | (ii) | 62 | 3e | |
| k | 42 | 2a | (nn) | 56 | 38 | |
| l | 45 | 2d | (rr) | 14 | 0e | |
| m | 16 | 10 | (tt) | 13 | 0d | |
| n | 11 | 0b | (yy) | 25 | 19 | |
| o | 23 | 17 | (ar) | 59 | 3b | |
| p | 9 | 9 | (aer) | 47 | 2f | |
| r | 39 | 27 | (ch) | 50 | 32 | |
| s | 55 | 37 | (ck) | 41 | 29 | |
| t | 17 | 11 | (ear) | 60 | 3c | |
| u | 15 | 0f | (eh) | 26 | 1a | |
| v | 35 | 23 | (er) | 51 | 33 | |
| w | 46 | 2e | (err) | 52 | 34 | |
| y | 49 | 31 | (ng) | 44 | 2c | |
| z | 43 | 2b | (or) | 58 | 3a | |
| (space) | 1 | 01 | (ou) | 22 | 16 | |
| (space) | 3 | 03 | (ouu) | 31 | 1f | |
| (space) | 4 | 04 | (ow) | 32 | 20 | |
| (oy) | 5 | 05 | ||||
| (sh) | 37 | 25 | ||||
| (th) | 29 | 1d | ||||
| (dth) | 18 | 12 | ||||
| (uh) | 30 | 1e | ||||
| (wh) | 48 | 30 | ||||
| (zh) | 33 | 21 |
Note that these only add up to 62 allophones. The two remaining ones are not implemented in the interpreter and are unlikely to be useful, although you can try them if you want - code 0 gives a 10ms pause, whilst code 54 (36h) gives an allophone practically identical to the (dth) allophone.
The speech signal is present as a high impedance output on the line lead as well as a UHF output. To record some speech, plug the line lead into the "MIC" input in your tape recorder, set it to "RECORD" and any speech produced by the unit will be recorded on tape. To output speech to an external amplifier, connect the line lead to the "AUX" or "TAPE" input on your amplifier, via a suitable connecting lead.
Before attempting to LOAD from a tape recorder, it is sensible to disable the keyvoices, as otherwise you will get a continuous "e" allophone outputted during the program load. This is due to the tape handling routines taking complete control of the Spectrum's CPU and preventing the MicroSpeech from updating its speech buffer. Upon completion of the LOAD, the "enter" keyvoice is completed normally.
If you leave the keyvoices on during SAVE, you will hear a keyvoice after the preliminary data block has been written to the tape recorder (the keyvoice you will hear is the one you press after the "press any key" prompt).
If you BEEP in the middle of an allophone, you will "draw out" the allophone and combine it with a musical tone, as the BEEP command prevents updates of the speech buffer for the duration of the tone. If you are very clever, you may be able to use this effect to make the MicroSpeech "sing".
The CURRAH MicroSpeech unit is fully compatible with the Sinclair ZX Microdrive and Interface I. Simply follow the setting up instructions supplied with Interface I, and plug the MicroSpeech unit into the port on the back of Interface I BEFORE plugging in the power.
Again, be sure to disable the keyvoices during all data transfer operations to prevent the "enter" keyvoice being drawn out.
Most games programs will run on the CURRAH MicroSpeech, and even if the program does not use the speech facility you will find the game is enhanced as the Spectrum's sound will come from your television speaker. Leading software houses are currently writing games software which will utilise the speech output - watch out for these as they become available.
The CURRAH MicroSpeech "mixes" sound into the UHF output of the Spectrum modulator, and consequently on some makes of TV set you may find some "patterning" where large areas of strong colour are being displayed. This is not due to a fault in your unit, but is an inevitable consequence of mixing sound into the UHF output; the technically-minded amongst you might be interested to know that the patterning is due to "sidebands" generated during the mixing process interfering with the colour information being transmitted to your television.