[fluid-dev] Purpose of dither?

[Miguel Lobo]

Hi list,

We have this code in fluid_synth.c:

#define DITHER_SIZE 48000
#define DITHER_CHANNELS 2

static float rand_table[DITHER_CHANNELS][DITHER_SIZE];

static void init_dither(void)
{
  float d, dp;
  int c, i;

  for (c = 0; c < DITHER_CHANNELS; c++) {
    dp = 0;
    for (i = 0; i < DITHER_SIZE-1; i++) {
      d = rand() / (float)RAND_MAX - 0.5f;
      rand_table[c][i] = d - dp;
      dp = d;
    }
    rand_table[c][DITHER_SIZE-1] = 0 - dp;
  }
}

This creates two tables of length 48000 with random floating point numbers between -1.0 and 1.0 (the probability distribution is not uniform).

And in the same file, in function fluid_synth_write_s16 (which converts the result of synthesis from floating point to signed 16-bit format):

    left_sample = roundi (left_in[cur] * 32766.0f + rand_table[0][di]);
    right_sample = roundi (right_in[cur] * 32766.0f + rand_table[1][di]);

    di++;
    if (di >= DITHER_SIZE) di = 0;

For each output sample, one of the previously generated random values is added to the result, which has now been scaled and varies between -
32766.0 and 32766.0.

I don't understand the purpose of all this.  The effect of adding the random values surely is increasing the noise in the output, and that by such a small amount that seems sure to be inaudible.

I'm no audio expert, however, so does anyone think that this code serves any useful purpose?

Regards,
Miguel