What do noise functions sound like?
Abigail Adegbiji • November 9, 2025
Lately I’ve been building a voxel game engine, which requires procedurally generated terrain. A nice approach is to break up the terrain into chunks and then sample a noise function to get the height values of the top layer voxels. This has led me to be curious about how different functions work, as well as how they sound like.
White noise
White noise is the simplest type of noise; it’s quite literally just a sequence of random values. A basic implementation might look like this:
float white()
{
static std::random_device rd;
static std::mt19937 gen(rd());
static std::uniform_real_distribution<float> dist(0.0f, 1.0f);
return dist(gen);
}
To actually hear the noise, the samples need to be written to an audio file. Of course, it’s be nice to listen to it, so we’ll write the noise to Here, we’ll just write to a wav file.
struct WavHeader
{
// riff header
uint8_t riff_chunk_id[4];
uint32_t total_file_size;
uint8_t riff_format[4];
// fmt chunk
uint8_t fmt_chunk_id[4];
uint32_t fmt_chunk_size;
uint16_t audio_format;
uint16_t num_channels;
uint32_t sample_rate;
uint32_t byte_rate;
uint16_t bytes_per_sample;
uint16_t bits_per_sample;
// data chunk
uint8_t data_chunk_id[4];
uint32_t data_chunk_size;
};
int main()
{
int seconds = 10;
uint32_t frequency = 44100; // samples per second
uint32_t num_samples = frequency * seconds;
int bytes_per_sample = sizeof(int16_t); // 16 bit pcm samples
uint32_t num_sample_bytes = num_samples * bytes_per_sample;
std::random_device device;
std::mt19937 engine(device());
std::uniform_int_distribution<int16_t> distribution(-32767, 32767);
// generate noise samples
std::vector<int16_t> samples;
samples.resize(num_samples);
for (size_t t = 0; t < samples.size(); t++) {
float noise = white();
samples[t] = (int16_t)(-32767.0f + noise * 65535.0f);
}
WavHeader header = {
.total_file_size = num_sample_bytes + 44,
.fmt_chunk_size = 16,
.audio_format = 1, // uncompressed samples
.num_channels = 1, // mono audio
.sample_rate = frequency,
.byte_rate = frequency * bytes_per_sample,
.bytes_per_sample = (uint16_t)bytes_per_sample,
.bits_per_sample = (uint16_t)(bytes_per_sample * 8),
.data_chunk_size = num_sample_bytes,
};
memcpy(header.riff_chunk_id, "RIFF", 4);
memcpy(header.riff_format, "WAVE", 4);
memcpy(header.fmt_chunk_id, "fmt ", 4);
memcpy(header.data_chunk_id, "data", 4);
std::ofstream output("output.wav", std::ios::binary | std::ios::out);
output.write((char*)&header, sizeof(WavHeader));
output.write((char*)samples.data(), samples.size() * sizeof(int16_t));
output.close();
}
The result is harsh and staticky because every sample is completely independent and random.
Brown noise
Brown noise (named after Brownian motion, not the color), takes a different approach. Instead of independent samples, brown noise uses a random walk process, where each sample is the previous sample plus a small, random step. This creates a correlation between consecutive values, resulting in a smoother sound.
float brown() {
const float decay = 0.999f; // prevents the value from drifting too far
static float value = 0;
static std::random_device rd;
static std::mt19937 gen(rd());
static std::uniform_real_distribution<float> dist(0.0f, 1.0f);
std::uniform_real_distribution<float> step_dist(-0.1f, 0.1f);
value = std::clamp(value * decay + step_dist(gen), 0.0f, 1.0f);
return value;
}
Brown noise sounds softer, deeper and more rumbling than white noise the random walk process naturally emphasizes lower frequencies.
Perlin noise
Neither white noise or brown works well for terrain generation. White noise creates completely chaotic landscapes while Brown noise drifts unpredictably without spatial coherence. What’s needed is coherent noise, where nearby points have similar values, creating smooth continuous variations. This is where algorithmns like Perlin Noise come in.
Perlin noise is a type of gradient noise that creates smooth, natural looking patterns. Imagine a grid where each intersection has an arrow pointing in a random direction (the gradient). To find the value at any point, look at the four surrounding grid corners, calculate how much each corner’s gradient influences the point, then interpolate those influences together.
This implementation requires a few helper functions and a hash function to generate random gradients.
struct vec2 { float x, y; };
inline float fade(float t) { return t * t * t * (t * (t * 6 - 15) + 10); }
inline float lerp(float a, float b, float t) { return (1 - t) * a + t * b; }
// Using xxHash for deterministic randomness
int hash(int x, int y)
{
int h = x * 3266489917 + y * 668265263;
h ^= h >> 15;
h *= 2246822519;
h ^= h >> 13;
h *= 3266489917;
h ^= h >> 16;
return h;
}
// Create a random gradient vector
vec2 gradient(int x, int y)
{
vec2 directions[] = {
vec2{1, 0}, vec2{-1, 0}, vec2{0, 1}, vec2{0, -1},
vec2{1, 1}, vec2{-1, 1}, vec2{1, -1}, vec2{-1, -1}
};
return directions[hash(x, y) & 7];
}
float perlin(float x, float y)
{
// get the grid cell the point's in and
// the direction of the point in that grid cell
int X = std::floor(x);
int Y = std::floor(y);
float dx = x - X;
float dy = y - Y;
// get the gradient vectors for each corner
vec2 gtl = gradient(X, Y);
vec2 gtr = gradient(X + 1, Y);
vec2 gbl = gradient(X, Y + 1);
vec2 gbr = gradient(X + 1, Y + 1);
// compute dot products to get the gradient values for each corner
float vtl = gtl.x * dx + gtl.y * dy;
float vtr = gtr.x * (dx - 1) + gtr.y * dy;
float vbl = gbl.x * dx + gbl.y * (dy - 1);
float vbr = gbr.x * (dx - 1) + gbr.y * (dy - 1);
// interpolate those values
float a = lerp(vtl, vtr, fade(dx));
float b = lerp(vbl, vbr, fade(dx));
float noise = lerp(a, b, fade(dy));
return noise * 0.7f + 0.5f;
}
Perlin noise works really well for terrain generation, but has some limitations. The square grid can cause subtle direcitonal artifacts, and it’s not the most efficient. Evaluating 4 grid corners (8 in 3D) and multiple interpolations per sample adds up.
Simplex noise
Simplex noise addresses these limitations. Instead of using a square grid, it uses a simplex grid, which has no directional bias and needs fewer corner evaluations (3 points in 2D instead of 4). Simplex noise first transforms the input coordinates from regular (x, y) space into a skewed grid space. Then it determines which simplex the point falls in, gets the contributions from each corner of the simplex then uses radial falloff, where each corner’s contribution fades out in a circular fashion.
float simplex(float x, float y)
{
// get the coordinate of the simplex cell the point's in
float skew = 0.5f * (sqrt(3.0f) - 1.0f);
int i = std::floor(x + (x + y) * skew);
int j = std::floor(y + (x + y) * skew);
// unskew back to (x, y) space
float unskew = (3.0f - sqrt(3.0f)) / 6.0f;
float x0 = x - (i - (i + j) * unskew);
float y0 = y - (j - (i + j) * unskew);
// determine the simplex the point's in (upper or lower triangle)
int i1 = x0 > y0 ? 1 : 0;
int j1 = x0 > y0 ? 0 : 1;
// get the gradients for the corner
vec2 g0 = gradient(i, j);
vec2 g1 = gradient(i + i1, j + j1);
vec2 g2 = gradient(i + 1, j + 1);
// get the corner coordinates
float x1 = x0 - i1 + unskew;
float y1 = y0 - j1 + unskew;
float x2 = x0 - 1.0f + 2.0f * unskew;
float y2 = y0 - 1.0f + 2.0f * unskew;
// sum the contributions from each corner
float sum = 0.0f;
float t0 = 0.5f - x0 * x0 - y0 * y0;
if (t0 > 0) sum += t0 * t0 * t0 * t0 * (g0.x * x0 + g0.y * y0);
float t1 = 0.5f - x1 * x1 - y1 * y1;
if (t1 > 0) sum += t1 * t1 * t1 * t1 * (g1.x * x1 + g1.y * y1);
float t2 = 0.5f - x2 * x2 - y2 * y2;
if (t2 > 0) sum += t2 * t2 * t2 * t2 * (g2.x * x2 + g2.y * y2);
// normalize to a range of 0 to 1
return 70.0f * sum * 0.5f + 0.5f;
}
The same gradient function from Perlin noise works here too.
Worley noise
Worley noise (also called Voronoi noise or Cellular noise) takes a completely different approach. Instead of interpolating between gradients, it measures distances to random feature points, creating cellular patterns with distinct boundaries between regions. The algorithm’s really simple. It divides the space into a grid, picks a random feature point in each grid cell, finds the nearest feature point for any query position and returns the distance to that point.
float worley(float x, float y)
{
int cell_x = std::floor(x);
int cell_y = std::floor(y);
float min_dist = std::numeric_limits<float>::max();
// check the current cell and the 9 neighboring cells
for (int dy = -1; dy <= 1; dy++) {
for (int dx = -1; dx <= 1; dx++) {
int neighbor_x = cell_x + dx;
int neighbor_y = cell_y + dy;
// generate a random feature point's position in this cell (0 to 1 range)
int h = hash(neighbor_x, neighbor_y);
float point_x = neighbor_x + ((h & 0xFFFF) / 65535.0f);
float point_y = neighbor_y + (((h >> 16) & 0xFFFF) / 65535.0f);
// get the distance to the point
float dist_x = x - point_x;
float dist_y = y - point_y;
float dist = sqrt(dist_x * dist_x + dist_y * dist_y);
min_dist = std::min(min_dist, dist);
}
}
return std::clamp(min_dist, 0.0f, 1.0f);
}
Fractal noise
All the noise functions show so far produce a single frequency (smooth at only one particular scale). But natural phenoma have detail at multiple scales. Fractal noise (also called fractional Brownian motion) solves this by simply layering multiple “octaves” of noise at different frequencies and amplitudes.
float fractal(float x, float y, int octaves)
{
float lacunarity = 2.0; // how much frequency decreases each octave
float persistence = 0.5; // how much amplitude decreases each octave
float frequency = 1.0f;
float amplitude = 1.0f;
float total = 0.0f;
float max_value = 0.0f;
// layer octaves of noise on top of each other
for (int i = 0; i < octaves; i++) {
// Can change the specific noise function used
total += simplex(x * frequency, y * frequency) * amplitude;
max_value += amplitude;
// increase frequency, decrease amplitude
frequency *= lacunarity;
amplitude *= persistence;
}
return total / max_value; // normalize to a range of 0 to 1
}
Fractal noise works with any base noise function, and creates this complex layered output.
Each noise function has its own character and use case. Perlin and Simplex noise provide smooth gradients. Worley creates cellular patterns and fractal noise adds detail at multiple scales. They also sound cool.
You can find the full source code here.