less_retarded_wiki/speech_synthesis.md

86 lines
2.5 KiB
Markdown
Raw Normal View History

2023-03-30 21:47:53 +02:00
# Speech Synthesis
TODO
## Example
This is a simple [C](c.md) program (using [float](float.md) for simplicity of demonstration) that creates basic vowel sounds using formant synthesis (run e.g. as `gcc -lm program.c && ./a.out | aplay`, 8000 Hz 8 bit audio is supposed):
```
#include <stdio.h>
#include <math.h>
double vowelParams[] = { // vocal tract shapes, can be found in literature
// formant1 formant2 width1 width2 amplitude1 amplitude2
850, 1650, 500, 500, 1, 0.2, // a
390, 2300, 500, 450, 1, 0.9, // e
240, 2500, 300, 500, 1, 0.5, // i
250, 600, 500, 400, 1, 0.9, // o
300, 400, 400, 400, 1, 1.0 // u
};
double tone(double t, double f) // tone of given frequency
{
return sin(f * t * 2 * M_PI);
}
/* simple linear ("triangle") function for modelling spectral shape
of one formant with given frequency location, width and amplitude */
double formant(double freq, double f, double w, double a)
{
double r = ((freq - f + w / 2) * 2 * a) / w;
if (freq > f)
r = -1 * (r - a) + a;
return r > 1 ? 1 : (r < 0 ? 0 : r);
}
/* gives one sample of speech, takes two formants as input, fundamental
frequency and possible offset of both formants (can model "bigger/smaller
head") */
double speech(double t, double fundamental, double offset,
double f1, double f2,
double w1, double w2,
double a1, double a2)
{
int harmonic = 1; // number of harmonic frequency
double r = 0;
/* now generate harmonics (multiples of fundamental frequency) as the source,
and multiply them by the envelope given by formants (no need to deal with
multiplication of spectra; as we're constructing the result from basic
2023-03-31 16:30:24 +02:00
frequencies, we can simply multiply each one directly): */
2023-03-30 21:47:53 +02:00
while (1)
{
double f = harmonic * fundamental;
double formant1 = formant(f,f1 + offset,w1,a1);
double formant2 = formant(f,f2 + offset,w2,a2);
// envelope = max(formant1,formant2)
r += (formant1 > formant2 ? formant1 : formant2) * 0.1 * tone(t,f);
if (f > 10000) // stop generating harmonics above 10000 Hz
break;
harmonic++;
}
return r > 1.0 ? 1.0 : (r < 0 ? 0 : r); // clamp between 0 and 1
}
int main(void)
{
for (int i = 0; i < 50000; ++i)
{
double t = ((double) i) / 8000.0;
double *vowel = vowelParams + ((i / 4000) % 5) * 6; // change vowels
putchar(128 + 127 *
speech(t,150,-100,vowel[0],vowel[1],vowel[2],vowel[3],vowel[4],vowel[5]));
}
return 0;
}
```