Elektor Vocoder (Part 1, January 1980) After all the articles describing the theory of vocoders, therere must be a lot of enthusiastic readers just itching to build one. This is just the sort of challenge that Elektor designers love: a lot of people want it, but nobody has corne up with a suitable design until now. Now, here it is at last! A 10-channel vocoder, designed in collaboration with Synton Electronics - an acknowledged specialist in the fied. The design offers good performance at a very reasonable cost. Ideal for musicians with a lot of enthusiasrn but insufficient funds to back it up! Certainly for those who would rather wield a soldering iron than a formula, the theory of vocoders has by now been discussed in more than adequate detail. Two years ago we discussed the 'hows and whys' and described the basic princiþles of a few cornmercially avaible vocoders. Last rnonth's article 'Vocoders' was intended as a brief recap of the history and technology of vocoders, and at the sarne tirne as a 'warming-up exercise for the construction project described this month. The difficulties associated with designing a vocoder were discussed at length. Obviousiy, these difficulties are even more apparent when the design is intended for home construction, as opposed to commercial production: the circuits must be absolutely reliable, and the effect of component tolerances must be reduced to a minimurn. Fortunately, the problems are not insurmountable, as we will see. Qne more time . . . We've explained what a vocoder is, often enough . . . 'We didn't really oughta repeat ourselves'. However. For those who are still unsure, in spite of all explanations given in earlier publications here is a brief definition: A vocoder is a 'box' with two inputs; one for a speech signal and one for a 'carrier' or 'replacement' signal ( in practice, this is usually some kind of 'music' signal). Inside the 'box', the speech characteristics are superirnposed on the carrier signal. A single output signal results. It contains all the characteristics (and intelligibility of the speech input, but the basic sound produced by the speaker (vibrations of vocal chords, resonances in the oral and nasal cavities) are replaced by those of the music signal. The result is something that sounds like music, but talks as well. How? This has been explained in many previous articles. However, in Specifications number of channels: 10 speech input sensitivity: adjustable 10 mV . . . 7.7 V impedance: 10 kOhm carrier input sensitivity: 770 mV impedance: l00 kOhm line output output level: 770 mV frequency range 30 . . . 16,000 Hz interest of providing a smooth transition to the block diagram and circuits that are to come, let us take a quick look at what is 'inside the box'. Most vocoders are so-called 'channel vocoders'. Other systems do exist (heterodyne-based, for instance) but these are so complex that they are rarely used in practice. The Elektor design is also a channel vocoder, so we will forget the other possibilities. Last month's article gave block diagrams that illustrate the basic principle. A quick look at the block diagram of the EIektor vocoder (figure 1 ) shows that it is almost identical. A channel vocoder consists of two main sections: the analyser and the synthesiser. These are very similar, both consisting mainly of an identical set of filters (two groups of ten in the Elektor vocoder). In the nalyser section, the filters are used to split the incoming speech signal into corresponding frequency bands. The output from each filter is rectified and passed through a low-pass filter; the total result is a set of varying DC voltages, each corresponding to the 'envelope' of the speech signal within that particular frequency band. The synthesiser section splits the 'carrier' signal into the same set of frequency bands. The output level in each band is varied by a voltage controlled amplifier (VCA) that is driven by one of the varying DC control voltages produced vy the analyser section. The result is that the amplitude 'envelope' of each frequency band in the speech signal is imposed on the corresponding frequency band of the carrier signal. The outputs of all VCAs are then summed, to produce the total output signal: basically, the tonal characteristics of the carrier signal with the articulation of the speech. Talking music, in other words. The Elektor vocoder After we had already done quite a bit of experimenting with vocoder circuits, we happened to come into contact with Synton Electronics -- the manufacturer of the well-known Syntovox vocoders. Some very profitable discussions with these specialists led to the circuit described here: a vocoder, designed specificaliy for home construction. The number of channels (frequency bands in the analyser and synthesiser sections) is limited to ten, for several good reasons. That number is adequate for good music reproduction and good 'speech' intelligibility; furthermore, it is a reasonable compromise between performance and price. Admittedly, a twenty-channel version sounds better more 'detailed'; but, in practice, the improvement is not often worth the vastly greater cost and compiexity required. Not only do you need twice as many filters: they must also be much 'steeper' /approximately 50 dB/octave and this requires careful design and expensive components. Usually, strict selection of components is neccessary for this type of filter - not a very feasible proposition for the average amateur. For a ten-channel vocoder, on the other hand, 24 dB/octave filters can be used. These are not nearly as complex and - even more important - quite reliable results can be obtained without having to resort to exotic components or test equipment. For that matter, reliability was an important factor in the design of the whole circuit - not only the filters. Wherever possible, the circuit is set up so that component tolerances and wiring will not affect the operation; furthermore, a larger number of adjustment points are included than is normal in professional equipment. By this means, good results can be obtained without the component selection normally required. Two features are deliberately omitted from the basic version: spectrum analysis and a voiced/unvoiced detector. The reason is obvious: although admittedly useful, these features are also expensive! However, the design does provide the option of adding them at a later date, and it is quite likely that we rvill be publishing suitable designs in the near future. For the present, however, we will do without. One little gimmick is included: ten LEDs, one for each channel, give an indication of the varying spectrum of the speech signal. Not that it has much practical use - but it doesn't cost anything, either. What does it all cost? An important consideration, for most people! From a quick look at figures 3 . . . 6 it is apparent that there are quite a few components in a vocoder. In plain language: it's crawling with opamps. To make matters worse, a lot of p.c. boards go into a unit of this kind - and they are not nearly as cheap as we would like. All in all, our estimate of the total cost works out at somewhere in the region of £100. A lot of money for a home construction project - but very cheap for a good vocoder, like this one. What's in the box? A block diagram of the vocoder is given in figure 1 . The upper half is the analyser section, and the lower half is the synthesiser. Let's take a look at the analyser first. The microphone signal is passed to a suitable preamplifier. Although not shown in the block diagram, the sensitivity of this input is adjustable over a wide range, so that it can also be used as a line input from an external microphone preamp. The preamplifier is followed by a buffer stage that includes bass cut, with a roll-off below approximately 30 Hz. The output from the buffer stage is fed to the filters that split it into frequency bands. Ten filters, corresponding to ten bands. Not all equal, however. Taken together, the filters cover the whole audio band, from about 30 Hz to 16 kHz, but the first filter (low-pass) and the tenth (high-pass) take care of a disproportionately large part of the spectrum. The low-pass filter covers the range from 30 Hz to 200 Hz; the high-pass is for everything above 4600 Hz. The central range, from 200 Hz to 4600 Hz, is most important for speech; it is divided into eight bands by the remaining filters. Each filter is followed by a precision rectifier and a low-pass filter. The latter is not shown, as such, in the blockdiagram - it is taken as an essential part of the rectifier stage. Obviously: for a vocoder, we are not interested in rapid fluctuations of the speech signal or remaining half- or full-wave rectified frequency components; what we want is the general level trend for each frequency band. The first stage in the synthesiser section is also a preamplifier, for the carrier signal this time. Once again, it is followed by a buffer stage - similar to the one in the analyser. From here, the signal is passed to the filters; these are identical to the first group. The output of each filter goes to a voltage-controlled amplifier (VCA). Each VCA receives its control voltage from the corresponding filter and rectifier in the analyser section. The output signals from all ten VCAs are summed; the total signal is passed to the output buffer stage. Finally, about all those dotted lines. In both the analyser and the equaliser section, the link between the input preamplifier and the following buffer stage is brought out, to create the possibility for adding a voiced/unvoiced detector at a later date. What happens is that both outputs and both inputs are brought out to a connector; from there, they run along a bus board to a further connector (intended for the detector); along the way, the copper tracks are deliberately bridged so that each amplifier output is connected to the corresponding buffer input. When a voiced/unvoiced detector is to be added, the bridge between the tracks must be broken. Furthermore, the connection between each rectifier output and the corresponding VCA control input is shown as a dotted line. These points are brought out to sockets on the front panel. This has the advantage that it is now possible to deliberately connect some or all of the outputs to the 'wrong' VCAs, for special effects. This will be discussed later, in greater detail, when it comes to 'using the vocoder'. For the moment, we are more interested in the electronic details of the various sections in the block diagram. Time for the circuits. The circuits A modular construction was choosen for the vocoder, as we will see later on. The various circuit sections are mounted on separate printed circuit boards. Twelve in all: one for the supply, one for the input amplifier and buffers plus the summing amplifier and output buffer, and one so-called filter unit board. This contains one complete section, as shown enclosed in dotted lines in figure 1 : two complete high-, low or bandpass filters with the associated rectifier and VCA. A more detailed block diagram of one filter unit is given in figure 2. Since the circuit of the complete vocoder is rather too extensive to swallow in one gulp - for that matter, it would be virtually impossible to print on a single magazine page - it is easier to deal with each circuit section separately. First the central building block of the vocoder: the filter unit. In particular, the band-pass filter version, as it appears eight times with only minor component value changes. The band-pass filter The circuit is given in figure 3. Those who may have felt that we were exaggerating when we said that the complete circuit was so extensive, should be having second thoughts by now. All of these components represent just one filter unit -- and there are ten of them in our vocoder. The band-pass version shown here is required eight times. Each one takes care of its own band in the total range (200 Hz . . . 4600 Hz), and this is obviously reflected in the component values. In particular, the values of capacitors C1 . . . C11. Table 1 gives the correct values for the bandpass filters BPF1 . . BPF8, with the resultant centre frequency of each filter. On talking a closer look at the circuit given in figure 3, it is not too difficult to recognise the various sections that make up the block diagram shown in figure 2. First, let's pin down the in- and outputs. Points 'a' and 'b' are the filter inputs for the analyser filter Ispeech and synthesiser filter (carrier), respectively: 'c' is the signal output - the output of the VCA, in other words. Point 'd' is the control voltage output from the rectifier (more properly, from the final low-pass filter) in the analyser: Vc,out: 'e' is the control voltage input, Vc,in for the VCA in the synthesiser. A1 and A2, with associated components, make up the band-pass filter in the analyser section. An identical configuration, using A5 and A7, does the same job in the synthesiser. The precision rectifier is constructed around A3 and A4; it is followed by the low-pass filter, using A9. Finally, A10 is the VCA. Admittedly, there are a few more opamps - but these will be discussed later. One thing is very obvious: there are a lot of opamps in this circuit. Not only in this one, for that matter - the whole vocoder is opamp-based. The main reason for this is to keep the circuit as simple as possible -- using transistors, it would really become messy . . . fortunately, the high-quality opamps that are readily available nowadays are quite suitable for audio work. Most of the opamps used in this filter unit are JFET-input types. There are four of them in a TL084. Another possibility is to use a 4741 - with the added advantage that its current consumption is lower. Both of these types have been used in previous Elektor designs, with good results, and availability should not be a problem. They cost about one pound each. A common-or-garden 741 is also used in the circuit and - for the VCA - an OTA, type CA3080. Quite familiar to Elektor readers ! The band-pass filters are of a fairly well-known type: in both sections, two so-called Rauch filters are connected in cascade. The slightly different component values for the first and second filter in each pair ensure that a slightly 'flattened' top is obtained for the total filter characteristic, instead of the sharp peak that a single filter would give. Each filter gives a slope of 12 dB/octave, so that two in cascade provide the desired 24 dB/oct. In passing, it is perhaps interesting to note that the slope of any properly-designed filter can be estimated by counting the 'active' capacitors and multiplying by 6. A single filter in this circuit contains two capacitors, making for 12 dB/octave. Back to the circuit. In the analyser section, the band-pass filter is followed by two opamps in a full-wave rectifier circuit (A3, A4, D1, D2) and an RC network (R30 and C9) to take care of the worst of the ripple. An active lowpass filter (A9) does the bulk of the smoothing. It is a good idea to tailor the low-pass filter to suit the frequency range selected by the preceding bandpass filter. For this reason, C9, C10 and C11 are given different values for each section, as listed in Table 1 . The no-signal DC component in the Vc,out control voltage should be zero, in the ideal case. For this reason, an offset adjustment (preset P1)/ is included for A9. The LED indication of the 'speech spectrum' that was mentioned earlier is obtained by using the same control voltage to drive a LED /D3) via a transistor (T1). In the synthesiser section, the first two opamps (A5 and A7) are used in the same filter configuration as that in the analyser. Then the VCA, for which an OTA (A10) is used. Since an OTA (Operational Transductance Amplifier) is basically a current-controlled amplifier - not voltage-controlled - a minor circuit extension is needed. The control voltage from the analyser section (Vc,in) is buffered (A6) and then fed to a voltage-to-current converter: A8 and T2. Basically, this is a voltagecontrolled current source; variations of the control voltage, Vc, are converted into variations in the bias current for the OTA (at pin 5 of A10). P4 is used to set a threshold value for this current -( LITTLE STRANGE HERE, BUT IT´S THE ORIGINAL!!!)the calibration procedure will be described later. The same applies for the calibration of P2, this adjustment is included to balance the input differential amplifier in the OTA - a necessary precaution to prevent the bias current variations breaking through to the output, in the absence of a 'carrier' signal. Low- and high-pass filters Figures 4 and 5 both bear a strong resemblance to the circuit given in figure 3. This is hardly surprising: the only real difference between the band-pass filter units (figure 3), the low-pass (figure 4) and the high-pass filter unit (figure 5) is the actual filter circuit. And even there, the difference is marginal. Both the low- and high-pass filters are standard variants on the well-known Sallen & Key filter. As before, two sections are connected in cascade to obtain a total filter slope of 24 dB/octave (four capacitors, remember?). The cut-off point for the low-pass filter is set at 200 Hz; for the high-pass filter, this is 4600 Hz. In- and output module The remainder of the vocoder proper is shown in figure 6: the in- and output circuits. These are all mounted on one p.c. board. For these sections, good signal-to-noise ratio and drive capability are extremely important. The 'ideal' opamp for this job is þe illustrious TDA1034 (or NE5534). If availability is a problem, an LF 357 can be used as a (temporary) replacement - although the signal-to-noise ratio will suffer. The speech input circuit is given in figure 6a. Opamp A31 is used as a very low-noise microphone preamp. The voltage gain can be set between x1 and x1000 for any input sensitivity between 10 mV ard 7,7V. The input impedance is roughly equal to 10 kOhm, and in practice microphones with almost any impedance can be used. A line input is also provided,'suitable for signals from an external microphone preamplifier; in this case, the gain is set to about xl2. The output from A31 is brought out, via the bus board, to a spare connector; from there, it comes back to the sensitivity control P13. As mentioned earlier, this is done to offer the possi- bility of adding a voiced/unvoiced detector at a later date. The sensitivity control is followed by a buffer /amplifier stage, A32. By adding C54 and C58, this stage also serves as an active rumble filter. Output 'a' from A32 is connected to all ten inputs 'a' on the filter units. Figure 6b is the 'carrier' input circuit. The sensitivity control, P14, is followed by an input preamplifier with a gain of approximately x10 (A33). As before, the signal then loops around the spare connector; finally, A34 is used as a combined buffer/amplifier/active bass-cut filter -- identical to the one in figure 6a. Output 'b' is again connected to all ten inputs 'b' on the filter unit boards. The outputs of all filter boards (point 'c' in figures 3, 4 and 5) are all connected to input 'c' in figure 6c: the input of the summing amplifier. The first stage (A35, an LM301 ) is followed by an output level control (P15) and an output buffer stage (A36). A TDA1034 is used for this final stage, for the same reasons given earlier (low noise and high output drive capability). The nominal (line) output level of the vocoder is approximately 700 mV; the output impedance is very low (a few ohms) due to the negative feedback: the effect of R 134 is cancelled (this resistor is included for stability and short-circuit protection). What's to come? The power supply circuit, printed circuit boards and parts lists are still outstanding. Then, of course, constructional details and calibration procedure. Quite a lot, all told, but we hope to squeeze it all in next month. What else? An article on 'using a vocoder' is scheduled, and there are plans for extending the LED indication - little more than a gimmick in the present design - so that the vocoder can be used as a simple spectrum analyser. A very useful extension. The further plans are rather more vague, but we certainly hope to do something about the voiced/unvoiced detector and associated noise generator in the not-too-distant future. One thing is for sure, we haven't heard the last of vocoders yet - not by a long chalk! Vocoder Elektors 57 & 58, January & February 1980. Page 1-28: several times in the text, A5 & A7 are transposed for A3 & A4 and vice-versa with respect to the circuit diagram, figure 3. The easiest remedy is to alter the circuit diagram. Page 1-29: again, several errors cropped up in the text. The 'line input' for the circuit of figure 6a is non-existent and the sensitivity control (P13) has moved! The circuit diagram itself (figure 6a) is correct apart from the addition of a 22 þF tantalum capacitor between R115 and 0 V, as shown in figure 6, page 2-19 (February issue). Page 2-16: capacitors C79 & C80, referred to on this page as being mounted on the bus board are in fact mounted on the input/output board. While we're on the subject of the bus board, many readers may have noticed that points 'g', 'h', 'i' and 'j' are incorrectly connected (bottom right of figure 9, page 2-23) : point 'g' should be connected to point 'h', while 'i' should be connect›d to 'j'. There is no need to worry, however, as the printed circuit board (EPS 80068-2) supplied by the EPS service is correct. Elektor Vocoder (Part 2, Febuary 1980) First, let's put one thing right. Last month, we stated that there were to be twelve printed circuit boards. Wrong: there are fourteen now. The wiring between the twelve original boards was getting so extensive that it was decided to plug them all into a so-called 'bus board' that runs along the back of the case. This board turned out to be so long that it had to be cut in two, for postal reasons. All other boards, with the exception of the power supply, are plugged into connectors on the bus board. This is a great help, both for construction and 'service' - so we hope no-one will complain about the two additional boards . . . Power supply Before getting to the p.c. board layouts, we must first provide the power supply circuit, as promised. As shown in figure 1, this circuit is so simple that it is hardly worth talking about. The symmetrical +/-15 V supply is obtained in the easiest possible way, using two integrated voltage regulators (IC19, IC20). The total current consumption is only 200 mA, so the 400 mA mains transformer will be more than adequate. Obviously, a larger transformer could be used, provided it fits in the case: future extensions, if and when they come, can then be powered from the same supply. For biasing the OTAs, a further symmetrical +/-5 V supply is also required. As shown in figure lb, these voltages are derived from the (stabilised) +/-15 V supply, by means of another pair of integrated voitage regulators (IC21 , IC22). The two tantalum electrolytics, C86 and C87, and the 100 n capacitors C84 and C85 are essential for this type of regulator: they suppress its annoying tendency to break into spontaneous oscillation. A printed circuit board for the supply is given in figure 2. To be more precise, it only accommodates the circuit shown in figure la; the +/-5 V supply (figure lb) is mounted on the bus board. A new feature We owe an explanation, although it is doubtful that many readers will have noticed it! ,lust before going to press last month, our esteemed 'boffins' came up with a small but very useful extension. It was included in the circuits for the high-pass filter and the input/output module (part 1 , figures 5 and 6) at the last minute, but we didn't quite get around to explaining it in the text - mainly owing to the fact that we were chasing around, trying to find out whether we were allowed to include it! The trouble was that our beautiful 'find' turned out to be patented - by Bode. We were still trying to find out how this effected us (fortunately, it doesn't) when the issue went to press, with the result that there were a few details in the circuits that remained completely unexplained in the text. This is common practice in industry, of course, but we feel that it is rather below-standard for a selfrespecting technical magazine. Our apologies! What extension? In figure 3, part of the high-pass filter is repeated. There's a potentiometer, P17, with a series resistor (R117). When we point out that the lower end of the series resistor is connected to the second input, 'K', of the summing amplifier (part 1 , figure 6), the basic idea may suddenly dawn. Some of the signal at the output of the high-pass filter (A11/A12) is taken off by P17 and added, without 'vocoding', to the final output. In this way, the lack of a voiced/unvoiced detector and associated noise generator can be camouflaged to some extent. More than 'some extent', in fact: the results can be surprisingly good! When the carrier signal is lacking in high-frequency content, there is not enough 'replacement signal' for the unvoiced 'hissing' sounds in speech (the 's', for instance). In this case, the high frequency components of the original speech signal can be added to the output signal; the correct 'blend' is set with P17. In many cases, this vastly improves the intelligibility of the vocoded signal. Provision is made for mounting the potentiometer, P17, on the p.c. board for the filter modules. The ground connection and that for the wiper ('f') are both at the edge of the board; the 'hot end' of the potentiometer is connected to a copper pad marked 'x' on the copper side of the board. Resistor R117 is mounted on the bus board. The connection from the lower end of this resistor to the input of the summing amplifier (points 'k') is included as a copper track on the bus board. Input/output and filter boards We can now do one of two things. Either repeat all the circuits already published last month, in part 1, or else ask you to dig out that January issue and refer to it as required. The latter option seems to be the most sensible. All right, so now we've got part 1 in front of us. A general block diagram of the filter units is given in figure 2, and complete circuits for the band-pass, low-pass and high-pass filter units in figures 3, 4 and 5, respectively. In the accompanying text, it was explained that a modular construction was to be used: one printed circuit board for each complete filter unit. No wild guess, this; in fact, our printed circuit board designer had already come up with a single, universal design for the filter board, suitable for all types of filter: low-pass, band-pass and high-pass. The layout of this universal filter board is given here, in figure 4. Figure 5 shows the component layouts, with accompanying parts lists, for mounting a band pass filter unit (figure 5a), low-pass filter (5b) and high-pass filter module (5c). The values for capacitors C1 . . . C11 in the eight band-pass filter units are listed in Table 1. This table was also included in part 1 , but it is repeated here with the rest of the parts lists. Observant readers may notice that the supplydecoupling capacitors (C73 . . . C76, 8 x C77 and 8 x C78, shown in figures 3, 4 and 5 in part 1) are missing in the layouts given in figure 5. Not to worry: they are included on the bus board. Then there's the board for the in- and output module (the circuit shown in part 1 , figure 6) . The copper and component layouts are given in figure 6. This p.c. board is exactly the same size as the filter unit board (70 x 168 mm). For that matter, the supply board (figure 2) is also the same size, even though it is not the intention at this time to mount it as a plug-in module. As before, the decoupling capacitors for the input/output module (C79 and C80) are mounted on the busboard. Now for a closer look at the boards. Mounting the components shouldn't be a problem - provided you don't get the various component layouts for the filter board mixed up. And don't forget the wire links; although they're not mentioned in the parts list, they do play an essential role. All connections to the boards are along the two ends. At one end, the connections associated with front-panel components; at the other end, the connector plug. On the filter boards, this means that the 'front' of the board contains the control voltage connections Uc out and Uc in (points d and e in the circuits), the LED output and the connections for the Ucc,in level control (8 x P3, P7, P11 ) . The 'rear' of the board contains all 'internal' connections: the speech and carrier inputs (points a and b), the vocoded output (point c), the supply connections and for special applications (to be described later) , a second set of control voltage connections (Uc,out and Uc,in). Similarly, on the input/output board, the front panel connections are at one end: input and output jacks with associated level controls (P13, P14, P15). The 'connector' end is for the suþply voltages and the internal in- and outputs a, b, c and k. This system means that each board can easily be built as a separate, plug-in module. A 21-pin connector is mounted on the 'inner' end of each of the filterunit boards and the input/output board (one suitable type is made by Siemens). The front panel is mounted at the other end it contains the control(s), jacks and LED. This construction is illustrated in figure 7: a sketch of a complete filterunit module. The small (3 mm) earphone jack shown are a good choice for the input connections. If the 'high-frequency blend' feature show n in figure 3 is to be added in the high pass filter unit, this will obviously call for a second potentiometer on its front panel. The input/output module also has a more densely populated front panel: it contains three potentiometers and three large-sized headphone jacks for the speech and carrier inputs and the vocoded output. Final assembly Now we come to the job of combining all the separate boards (or modules) into one complete 10-channel vocoder. The constructional block diagram (figure 8) illustrates the principle. It shows all the plug-in modules and the power supply; as can be seen, the bus board is a great help. Without it, the wiring would become rather messy. The letters a, b, c, d, e, and k, shown in figure 8, are also included on the various p.c. boards; they correspond to the indications in the circuits given in part 1. For simplicity, the supply is shown in figure 8 as a single board. In practice, as explained earlier, the +/-5 V supply is actually mounted on the bus board. P17 and R117 are also included in the block diagram; they are only required if the high-frequency blend option is to be added. Also shown in figure 8, enclosed in dotted lines, are the supply connections and two mysterious connection links. These refer to nine connections on the bus board, into which connector pins can be inserted. At a later date, they will provide an easy way to add a voiced/unvoiced detector with its associated noise generator. All supply voltages are available in this group, so that the unit can be powered from the main vocoder supply. The connection links between two pairs of contacts are actually those shown in the circuit of the input/output module (part 1 , figure 6) , at the outputs of A31 and A33. The links are already included as copper tracks on the board; when a voiced/unvoiced detector is to be added, these tracks are scratched away so that the speech and carrier signals run through this module. Having said so much about the bus board, it's time to take a look at it - or them, actually: as mentioned earlier, it is supplied in two sections that must be joined by means of wire links. Figure 9 shows the two p.c. boards and their component layouts. As can be seen, there was plenty of room between the eleven 21-pin 'female' connectors to mount the 5 V supply, ecoupling capacitors and one or two other odds and ends. One point has not been mentioned yet (nor shown in figure 8, to avoid confusion) : beside each connector, there are two connections for the Uc,in and Uc,out control voltages for each filter module. These are included with an eye to possible future extensions. For instance, in a complete system it may prove useful to route the control voltage interconnections through a plug-in matrix board, instead of using loose cables on the front panel. The various modules and the bus board are designed to fit neatly into a module case, as shown in figure 10. A standard 19 inch case can be used, with guide strips to hold the boards. This type of case is available from various manufacturers. The 19 inch width is just right for mounting the eleven modules at the spacing dictated by the bus board - no coincidence, this! The mains transformer and supply board can be mounted on the back plate, as shown in figure 10. A neat way to make the connections between the supply board and the bus board is by using so-called flat cable. For the various signai and control voltage in- and outputs, jack plugs are a good choice; the smaller (3 mm) type for all Uc,in and Uc,out connections and a larger version (6 mm) for the signal in- and outputs. Flexible cables with a small plug on each end can then be used to make all desired control voltage connections on the front panel. The mains switch, and an LED for power on/off indication, can be mounted on the front panel of the input/output unit. An alternative can be seen in figure 10: a potentiometer with built-in mains switch can be used for the output level potentiometer, P15. One word of warning, however: sometimes, the electrical screening between the switch and the potentiometer may prove inadequate - giving rise to an annoying hum. Alignment procedure We assume that everybody still has the original circuits, given in part 1 , to hand; in any event, we will be referring to them regularly . . . There are three preset potentiometers on each filter-unit module that must be correctly adjusted. This means that three separate adjustments must be performed for each board, as follows: 1. First the preset that sets a DC bias voltage for the inverting input of the OTA in each unit. In the eight band-pass filters, this is P2; on the low-pass filter board it is P10 and for the high-pass filter it is P6. The purpose of this adjustment is to ensure that the varying DC bias voltage, derived frorn the control voltage output of the analyser section when a speech input is present, cannot 'break through' to the 'vocoded' signal output. In simple terms: a signal present at point 'e' should not appear at output 'c'. This adjustment is carried out as follows: a. The Uc,out and Uc,in sockets on the front panel are interconnected by means of patch cords. b. All control voltage level potentiometers on the front panels (8 x P3, P7 and P11) are set to minimum, with the exception of the one on the module that is to be set up; that control is set to maximum. c. A steady noise signal is applied to the 'speech' input. One simple way to do this is to blow gently into the microphone. d. The bias potentiometer on that module (P2 for a band-pass filter, say) is adjusted for minimum output signal from the vocoder. If measuring equipment is available, a more precise alignment procedure can be considered. Instead of blowing into a microphone, a test signal can be applied direct to the Uc,in input of the module; a suitable test signal is a low-frequency sinewave (500 Hz or less), superimposed on a fixed DC voltage. The output signal from the vocoder can be observed on an oscilloscope, and the preset is adjusted for minimum LF output. In some of the modules, it may prove impossible to reduce the break-through to an acceptably low level. In this case, the OTA is almost certainly the culprit: in any batch there will always be a few that have too high a leakage from the control input to the output. The only solution is to replace them. 2. The next step is the preset in the voltage-to-current converter for the OTA: P4 in the band-pass filter units, P12 in the low-pass filter and P8 in the high-pass filter module. This adjustment is intended to set the initial point of the control characteristic to the same level for all modules. The procedure is as follows: a. A suitable test signal is applied to the 'carrier' input - white noise is a good choice. b. A very low DC voltage (approximately 200 mV) ,is applied to the Uc,in input of the module that is to be adjusted. This calibration voltage can be derived from the +5 V supply by means of a 25:1 attenuator (a 22 k resistor in series with 1 k, for instance). c. The control voltage level control on the front panel of the module (P3, P7 or P11) is set to maximum. d. The preset potentiometer (P4, P8 or P12) is now adjusted so that an output signal just appears at the main output. e. If the test voltage proves to be outside the adjustment range of one or more of the modules, the hole procedure can be repeated with a slightly higher or lower test voltage. 3. Finally, the easiest adjustment: P1, P5 and P9 in the band-pass, high-pass and low-pass modules, respectively. These presets determine the DC offset of the active low-pass filter that is the last stage in the analyser section of each module. With no (speech) input signal, each preset is adjusted for minimum Uc,out voltage of the corresponding module. In conclusion We've got an interesting photo for you, saved to the last. With a spectrum analyser and a lot of patience, we succeeded in plotting each of the filter characteristics separately and combining them in a single photo. The result of our efforts is shown in figure 11 . At the left in the display, the characteristic of one of the two identical) low-pass filters; this is followed by a neat procession of band-pass filter characteristics and,finally, the high pass filter. The minor differences in peak amplitude are caused by inavoidable component tolerances. Not that they have any real effect, in practice, since they can be compensated for by means of the control voltage level controls on the front panel. As can be seen, the filters provide a nicely regular division of the audio spectrum. Their Q is virtually identical, as is apparent from their equal band-pass 'widths' on this logarithmic frequency scale. This is by no means our last word on the subject of vocoders. Exactly what is to come, and when, has not yet been finally decided - so we won't make any promises. Anyway, for the time being all enthusiastic constructors should have plenty to do . . Elektor Vocoder (Part 3, February 1981) On the face of it, the detector may seem superfluous. However, when the block diagram of the complete vocoder in figure 1 is considered and the proposed additions are momentarily forgotten, their necessity will be readily apparent. In the upper section the speech signal is divided and split into control voltages to feed the VCA's in the synthesis section. The VCA's are thus provided with an input signal consisting of the carrier signal chopped into identical bits and pieces. Fair enough. In practice however, the synthesised result proves to be less satisfactory than expected. The fault lies with the carrier signal which is far from ideal. Most synthesised signals happen to be incomplete as far as their spectrum is concerned. This means that unvoiced sounds such as s, t, k and p do not come through very well, in fact they are often inaudible. The simple and effective remedy for this was the inclusion of the 'high frequency blend' provided by P17 shown in the dotted area in figure 1. Part of the 'high frequency' in the speech signal is taken from the high pass filter in the analysis section and is blended directly with the synthesised result. This is precisely what Harald Bode applies in his synthesiser. In practice this solves quite a few problems. For unvoiced signals to be properly synthesised, however, a circuit is required which can distinguish between the voiced and unvoiced sounds during analysis. Professionals call such a circuit a voiced/unvoiced detector and it is found in relatively few vocoders to date. The reason for this is largely due to the fact that the components reuired are fairly complex and therefore icrease the price of the vocoder considerably. Technically speaking, it is by no means easy to design and this of course also deters many manufacturers. When it is combined with a noise genertor a decent voiced/unvoiced detector is a great improvement on the blending trick mentioned earlier. The latter would not work, for instance, whenever speech is to be synthesised without an original speech signal. In other words, a microprocessor and a DA converter are unable to generate a complete, artificial speech spectrum. The detection system described here can however do this. It enables noise to be fed to all the synthesis filters in the vocoder whenever there are unvoiced sounds in the speech signal. With the aid of control voltages derived from the analysis seaion the required 'colour' noise can be produced. In addition, the detector is fast enough to provide a very true-to-life synthesis of the s, t, k and p sounds. How does it work? Whereas the practical construction is rather complicated, the block diagram of a voiced/unvoiced detector is fairly straight forward. Figure 1 shows the general principle. The speech signal is fed to a suitable detection system that can distinguish between the unvoiced and voiced sounds. This detector operates a switching circuit which interrupts the carrier signal in the event of unvoiced sounds and then substitutes it temporarily for the output signal of a noise generator. Clearly the detection system is at the heart of the matter, but the little block in the diagram hardly gives an indication of its function. What does it do exactly? Figure 2 illustrates the frequency ranges which the detector 'examines' before deciding whether the signal is voiced or unvoiced. The mere fact that there are many high frequencies in the speech signal does not mean that the speech signal is unvoiced at that moment. This assumption is totally incorrect, as the high frequencies measured may well be part of a complex signal with a fundamental frequency that is so low that it is a voiced signal after all. That is why the detector also checks the low frequency range (down to 600 Hz). If at that moment the range does not include signal, or if the signal is much smaller hat its high frequency counterpart, chances are the sound is indeed unvoiced. Thus, two elements must be incorporated in the detection system: a high pass filter with a cut-off frequency of about 2500 Hz and a low pass filter with a turnover point at about 600 Hz. The voiced/unvoiced detector The complete circuit diagram of the detector is given in figure 3. Points A, B and C of figures 3a and 3b are linked. Roughly speaking (there is a little more involved) the diagram in figure 3a constitutes the detection system and that in figure 3b the section drawn as a switch in the block diagram. Both circuits are mounted on a separate board. The noise generator is incorporated on a third board, but this will be dealt with later. First let us look at figure 3 in further detail. It can be seen that the speech signal derived from the vocoder initially reaches the buffer/amplifier A1 and is then split into two signals, each passing the filters mentioned above. The high pass is constructed around A2 and A3 and the low pass around A4 and A5. Their peak values are at 2500 Hz and 600 Hz, respectively. The two filter sections have a slope of 24 dB per octave to obtain the best possible separation. They are each followed by a rectifier (A6 and A8) and by a 12 dB/ octave smoothing filter (A7 and A9). The latter's turnover frequencies are around 300 Hz for the high pass system and 30 Hz for the low. The rectified and calibrated output signals are now fed to three amplifiers or comparators (A10, All, A12) followed by a number of logic gates. All that need be said about these is that they take care of the trigger signals that are required later on to feed the carrier or noise signal to the synthesis filters at the right moment. The 'voiced or unvoiced ?' decision mentioned with regard to figure 2 is taken by comparators A 10 . . . A12. Supposing an unvoiced signal arrives at the input, the output of A10 will become high and that of A11 will be low. In other words, the output of gate N1 will be low, that of N4 will be high and that of N11 will be low as well. In the case, where the signal is unvoiced, the output from the low pass filter will either be zero or at least smaller than that from the high pass filter. This means, the output of A11 will remain low causing that of gate N2 to be high and N 1 to be low. The final verdict will then be: unvoiced. If, on the other hand, the low pass filter produces a signal that is greater than that from the high pass, N1 will no longer be low and the outputs of A11 and A12 will both be high. The detector then decides: voiced. The other tri-state gates (N 10 . . . N 13) in figure 3b serve to switch off the detector if in the future it is to be controlled by means of a computer or microprocessor.The two LED indicators D15 (unvoiced) and D17 (voiced) display the state of the detector. Naturally, is considered superflous, the section around T4 and T5 can always be omitted. The switch indicated in the diagram actually consists of two VCAs, A16 and A17. These ensure that in the end either the carrier or the noise signal is fed to the synthesis filters. Further particulars Preset pots P1 and P2 preset the switch to voiced or unvoiced, as required. This can be done by alternately uttering 'A' and 'S' sounds in the microphone. Depending on the results, the sensitivity can be readjusted if necessary. P3 and P4 preset the trigger point of the comparators A10 and A12. This must be done simultaneously with P1 and P2. Switch Slab acts as a select switch for the voiced state. It has been added to enable musical instruments to be used as modulators as well. Whenever music is entered at the speech input, closing S1 will prevent a sudden noise from being fed to the filters at every high tone. Whatever the signal, the detector will always decide it is voiced. The inhibit input (Z) may be used to 'block' all the detector's decisions. Then of course the control inputs (V, X) must be provided with information. Again, this will come into effect once the unit can be controlled by a (micro)computer. OTAs A16 and All in the carrier/noise circuit need to be very carefully calibrated with the aid of P7 and P8. This must be achieved by a rectified signal at the control input (R66, R77). This method is spelled out in last year's March issue, vocoder constructors will no doubt remember the details. If the unit is not properly calibrated irritating click sounds will be produced when the detector is switched, which happens regularly in speech and singing. Figures 4 and 5 represent the track layout and component overlays of the voiced/unvoiced detector printed circuit boards. The detection circuit in figure 3a is incorporated on the board shown in figure 4, the remainder (figure 3b) being installed on the board in figure 5. The noise generator Figures 6 and 7 show the circuit diagram and the printed circuit board respectively of the noise generator. The noise generator is not only suitable for the vocoder but also for various other audio and acoustic measurements that demand a quality noise signal. The output can be switched from pink to white noise and vice versa. The unit consists of 7 commonly used ICs and a few passive components. There is no need to describe its operation in full detail here, as various noise generators have been published in Elektor recently. All of them have their pros and cons and this particular design may be considered a combination of them with the addition of a zero inhibit. This concerns pseudo random noise which is generated with the aid of a 31 bit shift register (IC3 . . . IC6). How this works was described in the January 1981 'Swinging Poster' article, where incidentally the same ICs were used. N1 and N2 together form a clock generator at a frequency of about 500 Hz. About 70 minutes are needed to run through a 31 shift register in all its states at this clock frequency. This will make the noise sufficiently 'random'. Diodes D1 . . . D31 combined with N3 provide the zero inhibit. As soon as the '000 . . . 0' state occurs, a ' 1 ' is entered in the shift register by way of N5. Gate N6 makes sure outputs 28 and 31 of the shift register are EXOR back coupled. Buffer N4 is followed by a filter which can be switched to pink or white noise, whichever is required. The white noise filter is a low pass filter at 23 kHz with an edge of 6 dB per octave. IC7 acts to amplifier the signal. The pink noise has to be slightly more amplified than the white, because its high frequencies have already been filtered out and so cannot contribute any further to it. P1 is used to equalise the output voltages for pink and white noise. The value indicated for the supply is based on that of the vocoder ( + 15 V) . However, the noise generator will work eqally well at + 12 V. The connections We are left with three new boards that have to be connected to the existing vocoder. From the block diagram in figure 1 it can be seen what the procedure basically involves. There are two possibilities: 1. Take an additional 'half bus board' (EPS 80068-2) . The three new boards are exactly the same size as the other vocoder boards and can all be provided with a similar connector. If a connector is mounted on all three, they can be inserted into the bus board straight away and this will then take care of the individual connections. That's all there is to it. The supply voltage(s) and points i, j and g obviously have to be derived from the vocoder bus board. How this is done is shown in figure 8. At the same time the additional half bus board provides a simple connection for the existing supply board belonging to the vocoder. This is an advantage, as there was no room for this on the original bus board. Now the supply board may be inserted into the additional half bus board and the connections remain as indicated in figure 8. Two more remarks: As illustrated in figure 1 , the existing connection between points i and j in the vocoder will have to be interrupted when the voiced/unvoiced detector is connected. The i-j connections will therefore have to be broken both on the 'old' and on the 'new' bus board. Finally, to avoid any misunderstanding: for the drawing of the connections in figure 8 the circuit board drawings of the old bus board were used. Be careful not to mount any components on the new half bus board, in spite of the indications in figure 8. 2. Don't use an additional half bus board - make the connections yourself. This will be necessary if the case is not wide enough for another three connecting boards so that the expansion boards will have to be mounted elsewhere in the case. The wiring required is shown in the diagram in figure 9. Again, of course, the i-j connection on the bus board will have to be broken. Final notes The 'computer' connections indicated in he diagram as: unvoiced in (V), unoiced out (W), voiced in (X), voiced out (Y) and inhibit (Z) are all situated on the front of the 'switch boards' given in figure 5. If required, these points can be led out quite simply with a connector. This will enable experimenters to control the unit by means of a computer without having to cope with complicated wiring problems. As the connection diagrams of figures 8 and 9 show, both the voiced/unvoiced detector and the noise generator can derive their supply voltage from the existing vocoder power supply. The current consumption of the three expansion boards adds up to about 100 mA for the +15 V voltage and to about 50 mA for the -15 V. Since the vocoder was issued with a 400 mA transformer, the extra consumption will by no means overload the circuit. People have told us that the - 15 V section of the original vocoder supply may encounter stability difficulties. his can be remedied by substituting C83 for a 2u2/25 V tantalum electrolytic capacitor and C85 for a 1 u/25 V type. Elektor Vocoder (Part 4, September 1980) Each channel in the vocoder contains three presets. Two of these are intended to eliminate leakage of the Voice and Carrier signals to the vocoder's output; the third sets the dynamic range of the voltage controi circuit (in the analyser section, where the audio signals are split up into small bands and are converted into DC control voltages). This is important if the vocoder is to respond to a wide range of input signal levels and reproduce the speech sounds as accurately as possible. In passing, it should be noted that this high 'responsiveness' may cause a disturbing side effect when the vocoder is used during live performances, where there is usually a high level of interference. In such cases the vocoder will analyse and synthesize the entire complex sound, producing an undesirable cacophony. Further on in this article, methods will be suggested to suppress these side-effects. For the moment, however, let us concentrate upon setting up the vocoder properly. The best way to start is to adjust potentiometers P1 , P5 and P9 in the band pass, high pass and low pass filters respectively. These presets compensate the output offsets of the filters that follow the rectifiers in the analyser section. To a large extent, this determines the vocoder's dynamic range. The offset should not be more than 5 mV. If this cannot be achieved it may be advisable to modify the offset compensation slightly, as shown in figure 1 . In the original design HA 4741 type opamps were used, as these have a smaller offset than the TL series. Unfortunately, they are also more difficult to obtain and more expensive. If all the Uout buses are now connected to the Uout buses, there is no danger of undesirable offset voltages turning on the OTAs in the synthesizer section (or cutting them off - if the offset is negative). The vocoder's dynamic behaviour is further determined by the following adjustment: the cut-off point of the OTAs. This can best be done with the aid of an oscillator and an oscilloscope or an AC millivoltmeter. The (sine wave) oscillator is connected to the carrier input and is tuned to each successive filter frequency in the synthesizer section. The signal voltage is set to about 10 V p-p, measured at pin 7 of A4, A14 and A24. The Uin potentiometer on the front panel is turned up fully and now the oscilloscope or millivoltmeter is used to check the output of A10, A20 or A30. The preset potentiometers P4, P8 and P12 are adjusted to the point where the output signal just stops decreasing (see figure 2). Finally, the leakage from control input to audio output of the OTAs must be reduced to a minimum. Usually, it will not be possible to eliminate this entirely - but it is worth while trying (even replacing the OTAs, if necessary), since break-through of the speech signal to the vocoder output seriously affects the overall performance. Figure 3 shows the measurement set-up; P2, P6 and P10 are adjusted for minimum break-through. Best results will be obtained when the leakage of the single phase rectified sine wave signal, applied to the speech inputs, is not greater than 5 mV p-p at the vocoder output. In practice, this will not be easy to achieve. It has been found that only 200 out of every 1 ,000 OTAs manage it! If an oscilloscope and an oscillator are available, it is a good idea to check the pass-band and gain of all the filters. Obviously, any deviation with respect to those particular aspects can lead to an undesirable colouring. If, however, good components are used (and mounted in the correct positions!), any error should be so small as to be negligible. How to use the vocoder Having set up the vocoder properly, the next question is what to do with it. Its most common application is as a 'voice processor'. A recent 'hit' in the charts is 'Funky Town' by Lipps Inc, in which the voices of two members of the group are transferred to the sound of a synthesizer. The introductory lyrics are difficult to understand (even for Americans!). One reason for this could be that the key chosen for the melody is rather high and, as our previous article on the vocoder stated, it is important that the frequency spectrum of the carrier signals overlap that of the speech input. If the carrier consists almost exclusively of high frequency components and the modulation signal (in this case the voice) is in a lower frequency range, only the higher harmonics of the voice will be superimposed on the carrier signal, as shown in figure 4. Furthermore, a woman's voice appears to be used as the modulation signal on this recording, with a formant range that is less suitable for he classical vocoder with a relatively small number of channels. Later on in 'Funky Town' the melody is played in lower key and a male voice sings the lyrics. The improved intelligibility is very noticeable! The Elektor vocoder has the advantage that it can offer a reasonable solution to the problem of non-overlapping frequency spectra. By connecting the voltage control outputs of the analyser to channels one or two pieces higher up in the spectrum instead of to the control input of the corresponding synthesizer channel, the significant spectral information is moved up, as it were, to range that encompasses the higher carrier frequencies. This technique, known as 'formant shift', will be dealt with in depth later on in this article. In addition to the vocoder's use as a voice processor there are many ways in which sounds can be superimposed on different kinds of carrier signals. The best way to get to know the vocoder is to systematically carry out experiments, using a microphone and a simple sawtooth or pulse generator. The microphone As far as the microphone is concerned, high quality type is best: if the modulation spectrum is free from coloration, the end product will also be good. Not everyone will be able to afford a highpriced microphone, of course, so a few suggestions on how to obtain good results with a reasonable quality microphone may prove useful. In the first place, it may prove useful to give the microphone pre-emphasis - in other words, emphasize certain frequencies, where necessary, or , .n-ate them. This is done by means of tone controls or with separate filters. One of the most important corrections to be made is to attenuate the low frequency range. It is difficult to, give precise figures for this, as it of course depends on the type of microphone used and also on the distance between the mouth and the microphone. The closer the microphone, the more low frequency components will reach the anaalyser, not to mention the sound of breathing and explosive consonants þ, k, etc.). Sometimes, depending on the high frequency spectrum of the carrier signals, it may be advisable to boost or attenuate the treble range. As a rule, a standard Baxandall tone control with a turnover frequency around 1 kHz is fine. The carrier Many sound sources rnay be used as carrier material, but a simple function generator with a control range between about 20 Hz and 1 kHz would be ideal for the first experiments. The most suitable wave forms to experiment with are triangle, square wave, sawtooth and pulseforms. Should such a generator not be available, you can always build one based on one of the many Elektor circuit designs. Monitoring the results The best way to judge the results is to use headphones. The system can also be used to drive a conventional audio system with loudspeakers, but headphones are preferable as they avoid acoustic feedback problems. A few simple examples When the microphone, generator and headphones are connected (figure 5) and everything is switched on, the first experiments may be carried out. If you don't want to fall back on sentences like 'Testing . . . one . . . two . . . three . . .' it is perhaps useful to have a text in front of you. Experience has taught us that not everyone possesses the 'gift of the gab' at such moments! The frequency of the generator is set at about 50-60 Hz, using a pulse waveform. The result will be a resonant, clear, synthesized voice. If the frequency remains unchanged, the result sounds like the 'Cylon effect'. Cylons are robot-like creatures from the American TV series and film: 'Battlestar Galactica'. A vocoder was in fact used to produce their robot voices. By raising the carrier frequency while continuing to speak, the synthesized voice can be made to change in pitch. It will become less intelligible once the frequency is above 500-600 Hz; this effect was mentioned earlier, when discussing the Funky Town recording. It should be clear that the pitch of the synthesized vocoder product depends exclusively on the carrier's pitch. The next test to be described will demonstrate this. The frequency is set to a low value, for instance 100 Hz, and now the pitch of the voice is changed by singing instead of speaking, or by producing other sound varying in pitch. You will notice that the resulting timbre will change, as if a band-pass filter were being used, but that the fundamental frequency will remain the same. This is because the generator is still set at a fixed frequency. Nevertheless, this is a source of regular misunderstandings. Witness the fact that the vocoder is often compared to a harmonizer or to a pitch shifter - equipment used to shift the fundamental frequency and the spectrum of speech or music. If the same good intelligibility is required at higher frequencies, 'formant shift' can be used. The Elektor vocoder is one of the few vocoders on the professional market that offers this interesting facility. Formant shift literally means shifting the intelligibility information to a higher or lower frequency range. By coupling the output voltages of the analyser to the control inputs of synthesizer filters which do not have the same Fo, the measured formants are transposed to another place in the spectrum. If, for example, the voice at the speech input is much lower than the fundamental frequency of the carrier signal, the result can be made more intelligible by shifting the formants to a higher carrier spectrum. The synthesized 'voice' will become clearer and at the same time assume an entirely different character. This phenomenon can be used with great success to produce 'funny' voices. The higher the analyser spectrum is moved up, the more the voice will sound like Donald Duck. If the analyser spectrum is transposed down, the speaker will sound as if he suffers from the proverbial hot potato. Quite a different way to manipulate the formants is 'formant inversion'. To obtain this effect the analyser and synthesizer channels are cross-coupled. Not surprisingly, the result will be practically unintelligible. All transient sounds, such as K, P, T and hissing sounds will be superimposed on the low end of the carrier spectrum, whereas the low frequency information in the speech signal will control the high end of the carrier spectrum. Furthermore, of course, the formants will be thoroughly mixed. A good example of this is the '0' sound which comes out as a 'U'. In spite of the fact that the result is virtually unintelligible, this effect can be useful when making (complex) musical sounds. This is illustrated in figure 6. The results obtained so far through speech synthesis will all sound robotlike. In the first place, this is due to the pulse signal used as a carrier: it contains a lot of higher harmonics, creating a slightly grating, 'mechanical' sound. If a sawtooth is used instead of a pulse shaped signal as a carrier, the result will be softer. This illustrates that the carrier's complexity affects the timbre. To attenuate the robot sound further there are all sorts of other tricks. By modulating the carrier signal, for instance with a low frequency sinewave or triangular signal, a much more life-like 'human' sound is produced. Other modulation effects may involve a low frequency random signal or, even better, a control signai that is derived from the fundamental frequency of the original speech. This can be simulated by tuning the generator to the voice pitch and then adjusting it by hand to follow the inflections. When an accurate frequency/voltage converter ('pitch extractor') is used a very natural sounding voice can be synthesised, which shows that the intonation of the voice is a very essential part of human speech. A few suggestions to obtain carrier modulation are given in figure 7. Unvoiced consonants Up to now, the unvoiced consonants (S, SH, SK, SY, K, T, P, F , etc.) have been neglected. These cannot be successfully reproduced by only using a sawtooth or pulse as a carrier. To synthesize unvoiced consonants, a detection system is required with the aid of which noise can be added to the carrier signal at the right moment. Since the Elektor vocoder does not (yet) possess that Voiced/Unvoiced detector, another trick will have to be used for the moment. A very clever expedient was developed by Harald Bode, vocoder manufacturer, and he has now taken out a patent for it. Bode constructed a sort of 'bypass' circuit for high frequencies derived from the analyser section. In the case of the Elektor vocoder this has been provided by means of potentiometer P17 on the highpass filter. This contains the high frequency range of the speech spectrum where most unvoiced sounds are produced. By adding this signal directly to the output, a reasonably complete 'speech' signal may be obtained. Nevertheless, it is worthwhile to listen to the unvoiced sounds as they are reproduced when pulse or sawtooth waves form the carrier signal. By producing hissing and 'plop' noises in the microphone while switching the generator from triangle to squarewave to sawtooth to pulse waveshapes, you can hear how important it is to have a wide carrier spectrum for unvoiced sounds. Using a triangular wave, which only has evern harmonics, the result will be very poor, whereas the pulse which contains all the harmonics will produce something remotely like an S or an F. Whistling into the microphone with a fixed pulse frequency as a carrier will also show how much high frequency energy it possesses. The vocoder for musicians The experiments just carried out may seem a little too simple, but they emphasize the basic operation of the vocoder. Once the user really feels he understands exactly what is happening, the variety of applications will only be limited by his imagination. When used for musical applications, the vocoder will be restricted to keyboard and string instruments. After all, a saxophone player can hardly be expected to blow and talk and sing all at the same time! Guitar and bass guitar players will discover that more often than not the dynamic range of their instrument will not be sufficiently wide to produce intelligible or clearly articulated sounds. Depending on the effect that they whish to achieve, it may be advisable to connect an effects box between their instrument and the vocoder carrier input, with which additional high frequency components may be added to the original sound. Examples of such devices are phasers, flangers, boosters, distorters, fuzzers, frequency doublers, etc. It may also be interesting to connect the guitar to the speech input of the vocoder, while using an organ, string quartet or synthesizer as the carrier signal. This of course requires strict coordination between the various players. Chords or a melody will be played on the keyboard instrument, whereas the guitar is used to play a melody or a rhythmic pattern - preferably monophonic, so no chords. The newly generated sounds will have the envelope shape and some of the spectral characteristics of the guitar. Many other musical instruments may of course be equally well combined. For electronic pianos the same applies as for the guitar. Here too, the use of some kind of effects device is recommended. Organists and synthesizer players have a much easier time. A nice effect which can be produced on most keyboard instruments is the bass effect, by making explosive noises with the mouth in the microphone and letting them decay. Wind instruments like the tuba, trombone, etc. can be imitated with a little practice. Electronic synthesizers, like the Elektor Formant, offer an extremely wide range of possibilities. Apart from generating carrier sounds, the synthesizer can also be used to produce signals to control the vocoder synthesizer inputs directly, and the analyser outputs of the vocoder can be used to control numerous units in the modular synthesizer. The vocoder at live performances When performing with the vocoder on stage during a concert, a few aspects need to be treated with care. There are basically two characteristics in the vocoder, which could turn the performance into an absolute catastrophe. In the first place its sensitivity or 'responsiveness' which was mentioned earlier. Like so many devices, the Great Compromise will have to be sought. Providing the vocoder with a wide dynamic range may create chaos in noisy surroundings. This is because the vocoder makes no distinction between what it hears and what it is supposed to hear. ('Not in front of the vocoder!') Everything that enters the analyzer is processed in the usual fashion and appears synthesized at the output of the equipment and those of you who have experienced the result know what a terrible din that can be! The only suitable methods to suppress such sensitivity to undesirable noises is to use a highly directional microphone which is spoken into from as short a distance as possible or to use two microphones in antiphase. The latter method is illustrated in figure 8. When two (identical) microphones are used in this way it is important to speak or sing in front of one of them at as short a distance as possible. A plop cap and a bass roll-off filter are indispensable. Another advantage of this method is that acoustic feedback may be noticeably reduced. Feedback sensitivity happens to be another drawback of the vocoder, as a result of the phase shifts in ranges where the syntheser filters overlap. The vocoder in the studio The above-mentioned precaution to curb nasty side effects are of course less important in recording studios and may even be totally unnecessary. The vocoder is an instrument which is highly suitable for use in the studio, provided that a few details are taken into account - particularly when dealing with existing recordings. The voeoder is not a miracle machine with a 'talent button' or a 'success filter', but an instrument which one must learn to use, preferably in the initial stages of a musical production, where required. If 'vocoding' is postponed until all the material is recorded on the various tracks of a multi track recorder, there is a chance that the material may not be spectrally wholly suitable and that the synchronisation between the Voice and Carrier signals may not be sufficient. The problem in the sound studio is often that 'time is money' and so a producer will sometimes get a little impatient if the vocoder does not obtain astounding results at first bat. \/ocoding is then postponed until the final mix-down stage, where it is often much more difficult to obtain the desired effect. Fortunately, more and more sound technicians seem to understand that the vocoder needs to be played, like any other instrument, and that learning to play may take some time. Finally, figure 9 provides a few examples in which the vocoder can play an interesting part, especially if more voltage control equipment is available. Figure 10 gives a few suggestions for peripheral devices to make the vocoder more versatile. The voiced/unvoiced detector, in particular, is scheduled for publication in the near future.