Chatting Robot
by Paul Batchelor

"Chatting Robot", also known as "computer talking on the phone with his Mother", is a patch centered around a rudimentary vowel synthesis technique. This technique involves an approximate glottal excitation signal sent through three bandpass filters tuned at specific formant frequencies.

Table generation

The tables are used to store parameters for the three bandpass filters. Tables are used to interpolate between three different states, each corresponding to a different vowel sound:

• position 0 is an 'o' sound
• position 1 is an 'a' sound
• position 2 is an 'e' sound

The interpolation between the states allows the voice to smoothly morph between vowels.

Tables f1, f2, and f3 correspond to the first, second, and third bandfilters, respectively.

 # o a e
 'f1' '350 600 400' gen_vals
 'f2' '600 1040 1620' gen_vals
 'f3' '2400 2250 2400' gen_vals

Tables g1, g2, and g3 correspond to the gains of the bandpass filters.

 'g1' '1 1 1' gen_vals
 'g2' '0.28184 0.4468 0.251' gen_vals
 'g3' '0.0891 0.354 0.354' gen_vals

Tables bw1, bw2, and bw3 correspond to the bandwidths of the bandpass filters.

 'bw1' '40 60 40' gen_vals
 'bw2' '80 70 80' gen_vals
 'bw3' '100 110 100' gen_vals

Jitter

The jitter signal is in charge of morphing between the vowel sounds. The core of the jitter signal is randi, who range and between 0 and 2, and whose speed is modulated by a sine wave between 2 and 10.

The output signal is then copied to p-register 0.

 0 2
 0.1 1 sine 2 10 biscale
 randi 0 pset

Glottis

To simulate the glottis, a bandlimited pulse oscillator is used with a now pulse width, or duty cycle.

The frequency of the glottis is randomly modulated with another randi jitter signal in the range of 110 and 200 Hz. The rate of change is modulated with a sine wave in the range of 2 and 8 Hz.

 110 200 0.8 1 sine 2 8 biscale randi 

Additional deviation is added using jitter.

 3 20 30 jitter + 

Finally, the rest of square, with an amplitude of 0.2 and a pulse width of 0.1.

 0.2 0.1 square
 dup

Resonant Cavities

The glottal signal is fed through three bandpass filters in series to generate vowel sounds.

As mentioned in the previous section, parameters are linearly interpolated between three states in order to morph between vowel sounds. The interpolation is done using tabread.

 0 p 0 0 0 'g1' tabread *
 0 p 0 0 0 'f1' tabread
 0 p 0 0 0 'bw1' tabread
 butbp

and here are the other two filters...

 swap dup
 0 p 0 0 0 'g2' tabread *
 0 p 0 0 0 'f2' tabread
 0 p 0 0 0 'bw2' tabread
 butbp
 swap
 0 p 0 0 0 'g3' tabread *
 0 p 0 0 0 'f3' tabread
 0 p 0 0 0 'bw3' tabread
 butbp
 + +

Words

The signal generated up to this point is pretty good, but it is a endless babble with no breaks. It would be good to add some pockets of silence to add space and create words and phrases. This effecting is created using a makeshifte envelope generator, built from a maygate and a portamento filter. The maygate is being fed by a dmetro, sending off a tick every 400 milliseconds.

 0.4 dmetro 0.5 maygate 0.01 port * 2.0 *

Effects

Finally, the resulting signal is sent into zitareverb.

 dup
 60 500 3.0 1.0 6000
 315 6
 2000 0
 0.2 2 zitarev drop