Spaces:
Sleeping
Sleeping
new rain simulation method
Browse files- README.md +7 -3
- app.py +412 -406
- requirements.txt +1 -0
README.md
CHANGED
|
@@ -10,15 +10,19 @@ pinned: false
|
|
| 10 |
license: mit
|
| 11 |
---
|
| 12 |
|
| 13 |
-
# Granular Synthesis
|
| 14 |
|
| 15 |
Educational demo of granular synthesis with two modes.
|
| 16 |
|
| 17 |
## Rain Simulation
|
| 18 |
|
| 19 |
-
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
Three layers:
|
|
|
|
| 22 |
|
| 23 |
## Tonal Granular
|
| 24 |
|
|
|
|
| 10 |
license: mit
|
| 11 |
---
|
| 12 |
|
| 13 |
+
# Granular Synthesis
|
| 14 |
|
| 15 |
Educational demo of granular synthesis with two modes.
|
| 16 |
|
| 17 |
## Rain Simulation
|
| 18 |
|
| 19 |
+
Spectral-domain rain synthesis. White noise is sculpted in the frequency domain
|
| 20 |
+
to match the spectral profile of real rainfall (energy concentrated 2 to 12 kHz,
|
| 21 |
+
with slope varying by intensity). Each STFT frame is a "grain" whose spectrum
|
| 22 |
+
is shaped by the rain profile, then overlap-added into the output.
|
| 23 |
|
| 24 |
+
Three layers: continuous spectral wash, sparse transient drops (scipy bandpass filtered),
|
| 25 |
+
and optional thunder rumble.
|
| 26 |
|
| 27 |
## Tonal Granular
|
| 28 |
|
app.py
CHANGED
|
@@ -1,287 +1,369 @@
|
|
| 1 |
"""
|
| 2 |
-
Granular Synthesis Demo
|
| 3 |
-
===============================================
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
"""
|
| 14 |
|
| 15 |
import numpy as np
|
|
|
|
|
|
|
| 16 |
import gradio as gr
|
| 17 |
|
| 18 |
# ---------------------------------------------------------------------------
|
| 19 |
# Constants
|
| 20 |
# ---------------------------------------------------------------------------
|
| 21 |
SR = 44100
|
| 22 |
-
DURATION =
|
| 23 |
|
| 24 |
# ---------------------------------------------------------------------------
|
| 25 |
-
#
|
| 26 |
# ---------------------------------------------------------------------------
|
| 27 |
|
| 28 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
"""
|
| 30 |
-
|
|
|
|
| 31 |
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
A biquad gives a proper -12 dB/octave roll-off with a tunable
|
| 36 |
-
cutoff frequency, which is essential for shaping rain timbre.
|
| 37 |
|
| 38 |
-
|
|
|
|
| 39 |
"""
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|
| 50 |
# Normalize
|
| 51 |
-
|
| 52 |
-
a = np.array([1.0, a1 / a0, a2 / a0])
|
| 53 |
-
|
| 54 |
-
# Apply filter (direct form II transposed)
|
| 55 |
-
out = np.zeros_like(signal)
|
| 56 |
-
z1, z2 = 0.0, 0.0
|
| 57 |
-
for i in range(len(signal)):
|
| 58 |
-
x = signal[i]
|
| 59 |
-
y = b[0] * x + z1
|
| 60 |
-
z1 = b[1] * x - a[1] * y + z2
|
| 61 |
-
z2 = b[2] * x - a[2] * y
|
| 62 |
-
out[i] = y
|
| 63 |
-
return out
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
def brownian_noise(n_samples: int) -> np.ndarray:
|
| 67 |
-
"""
|
| 68 |
-
Brownian (red) noise = integrated white noise.
|
| 69 |
|
| 70 |
-
|
| 71 |
-
like a deep, smooth rumble — much closer to steady rain wash
|
| 72 |
-
than white noise (which sounds like static/frying).
|
| 73 |
-
"""
|
| 74 |
-
white = np.random.randn(n_samples) * 0.02
|
| 75 |
-
brown = np.cumsum(white)
|
| 76 |
-
# Remove DC drift and normalize
|
| 77 |
-
brown -= np.mean(brown)
|
| 78 |
-
peak = np.max(np.abs(brown))
|
| 79 |
-
if peak > 0:
|
| 80 |
-
brown /= peak
|
| 81 |
-
return brown
|
| 82 |
|
| 83 |
|
| 84 |
# ---------------------------------------------------------------------------
|
| 85 |
-
#
|
| 86 |
# ---------------------------------------------------------------------------
|
| 87 |
|
| 88 |
-
def
|
| 89 |
-
size_ms: float,
|
| 90 |
-
cutoff_hz: float,
|
| 91 |
-
resonance: float = 1.0,
|
| 92 |
-
) -> np.ndarray:
|
| 93 |
"""
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
A real raindrop sound has three phases:
|
| 97 |
-
1. IMPACT — very short broadband transient (< 1 ms)
|
| 98 |
-
2. BODY — the surface resonates briefly (metal rings, glass taps)
|
| 99 |
-
3. TAIL — fast exponential decay into silence
|
| 100 |
|
| 101 |
-
We
|
| 102 |
-
- A two-stage envelope (sharp attack + tunable decay)
|
| 103 |
-
- A biquad low-pass at a cutoff that varies with drop "size"
|
| 104 |
-
- Resonance (Q factor) that models the surface material
|
| 105 |
"""
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
# --- Two-stage envelope: fast attack, variable decay ---
|
| 110 |
-
# The attack is near-instant (first 5% of grain).
|
| 111 |
-
# The decay rate determines how "ringy" vs "dead" the surface is.
|
| 112 |
-
attack = np.minimum(t / 0.02, 1.0) # ramp up in first 2% of grain
|
| 113 |
-
decay = np.exp(-6.0 * t) # smooth exponential tail
|
| 114 |
-
envelope = attack * decay
|
| 115 |
-
|
| 116 |
-
# --- Noise source ---
|
| 117 |
-
# WHY noise and not a sine? Water impact is chaotic — it excites
|
| 118 |
-
# all frequencies at once. The filter then shapes the spectrum.
|
| 119 |
-
noise = np.random.randn(n)
|
| 120 |
-
|
| 121 |
-
# --- Apply envelope BEFORE filtering ---
|
| 122 |
-
# This way the filter's transient response adds a natural "ring"
|
| 123 |
-
# to the attack, which sounds like a surface being excited.
|
| 124 |
-
shaped = noise * envelope
|
| 125 |
-
|
| 126 |
-
# --- Biquad low-pass with resonance ---
|
| 127 |
-
# cutoff_hz controls brightness (glass=high, soil=low)
|
| 128 |
-
# resonance (Q) controls how much the surface "rings"
|
| 129 |
-
Q = 0.707 + resonance * 2.0 # 0.707=flat, higher=resonant peak
|
| 130 |
-
filtered = biquad_lowpass(shaped, cutoff_hz, Q=Q)
|
| 131 |
-
|
| 132 |
-
# Normalize grain
|
| 133 |
-
peak = np.max(np.abs(filtered))
|
| 134 |
-
if peak > 0:
|
| 135 |
-
filtered /= peak
|
| 136 |
|
| 137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
|
| 140 |
# ---------------------------------------------------------------------------
|
| 141 |
-
#
|
| 142 |
# ---------------------------------------------------------------------------
|
| 143 |
|
| 144 |
def synthesize_rain(
|
| 145 |
rain_type: str,
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
surface_brightness: float,
|
| 150 |
-
surface_resonance: float,
|
| 151 |
stereo_width: float,
|
|
|
|
|
|
|
| 152 |
) -> np.ndarray:
|
| 153 |
"""
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
Architecture:
|
| 157 |
-
Layer 1 — Individual drops (granular scatter)
|
| 158 |
-
Layer 2 — Continuous wash (filtered brownian noise)
|
| 159 |
-
Layer 3 — Thunder rumble (optional, low sine cluster)
|
| 160 |
|
| 161 |
-
|
| 162 |
-
|
|
|
|
| 163 |
"""
|
| 164 |
n_out = int(DURATION * SR)
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
#
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
#
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
# Thunder envelope: slow build, long tail
|
| 250 |
-
env = np.exp(-1.2 * t) * (1 - np.exp(-8 * t))
|
| 251 |
-
rumble *= env * 0.35
|
| 252 |
-
|
| 253 |
-
end = min(th_pos + th_len, n_out)
|
| 254 |
-
seg = rumble[:end - th_pos]
|
| 255 |
-
# Thunder is roughly centered in stereo
|
| 256 |
-
output[th_pos:end, 0] += seg * 0.8
|
| 257 |
-
output[th_pos:end, 1] += seg * 0.8
|
| 258 |
-
|
| 259 |
-
# --- Final normalization ---
|
| 260 |
-
peak = np.max(np.abs(output))
|
| 261 |
if peak > 0:
|
| 262 |
-
|
| 263 |
|
| 264 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 265 |
|
| 266 |
|
| 267 |
# ---------------------------------------------------------------------------
|
| 268 |
-
# Tonal granular engine (unchanged
|
| 269 |
# ---------------------------------------------------------------------------
|
| 270 |
|
| 271 |
def make_tonal_source(freq: float = 220.0, duration: float = 2.0) -> np.ndarray:
|
| 272 |
t = np.linspace(0, duration, int(SR * duration), endpoint=False)
|
| 273 |
-
|
| 274 |
for k in range(1, 8):
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
return
|
| 278 |
|
| 279 |
|
| 280 |
def granular_synthesize(source, grain_size_ms, density, randomness, pitch_shift):
|
| 281 |
grain_samples = max(int((grain_size_ms / 1000.0) * SR), 64)
|
| 282 |
window = np.hanning(grain_samples)
|
| 283 |
hop = max(int(grain_samples / density), 1)
|
| 284 |
-
|
| 285 |
n_out = int(DURATION * SR)
|
| 286 |
output = np.zeros(n_out, dtype=np.float64)
|
| 287 |
|
|
@@ -302,7 +384,6 @@ def granular_synthesize(source, grain_size_ms, density, randomness, pitch_shift)
|
|
| 302 |
rand_pos = np.random.randint(0, max(src_len - grain_samples, 1))
|
| 303 |
start = int(seq_pos * (1 - randomness) + rand_pos * randomness)
|
| 304 |
start = np.clip(start, 0, src_len - grain_samples)
|
| 305 |
-
|
| 306 |
grain = pitched[start: start + grain_samples] * window
|
| 307 |
out_pos = i * hop
|
| 308 |
if out_pos + grain_samples > n_out:
|
|
@@ -317,11 +398,11 @@ def granular_synthesize(source, grain_size_ms, density, randomness, pitch_shift)
|
|
| 317 |
|
| 318 |
|
| 319 |
# ---------------------------------------------------------------------------
|
| 320 |
-
#
|
| 321 |
# ---------------------------------------------------------------------------
|
| 322 |
|
| 323 |
-
def cb_rain(rain_type,
|
| 324 |
-
audio = synthesize_rain(rain_type,
|
| 325 |
return (SR, audio.astype(np.float32))
|
| 326 |
|
| 327 |
|
|
@@ -332,173 +413,98 @@ def cb_tonal(grain_size, density, randomness, pitch_shift, freq):
|
|
| 332 |
|
| 333 |
|
| 334 |
# ---------------------------------------------------------------------------
|
| 335 |
-
#
|
| 336 |
# ---------------------------------------------------------------------------
|
| 337 |
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
.
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 349 |
|
| 350 |
-
|
| 351 |
footer { display: none !important; }
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
h1, h2, h3 {
|
| 355 |
-
font-family: 'Inter', sans-serif !important;
|
| 356 |
-
font-weight: 600 !important;
|
| 357 |
-
color: #e6edf3 !important;
|
| 358 |
-
letter-spacing: -0.02em !important;
|
| 359 |
-
}
|
| 360 |
-
h1 { font-size: 1.75rem !important; }
|
| 361 |
-
h3 { font-size: 1.05rem !important; color: #8b949e !important; }
|
| 362 |
-
|
| 363 |
-
p, span, label, .gr-prose {
|
| 364 |
-
font-family: 'Inter', sans-serif !important;
|
| 365 |
-
color: #c9d1d9 !important;
|
| 366 |
-
font-size: 0.9rem !important;
|
| 367 |
-
line-height: 1.6 !important;
|
| 368 |
-
}
|
| 369 |
-
|
| 370 |
-
/* ── Panels and cards ── */
|
| 371 |
-
.gr-panel, .gr-box, .gr-form, .gr-block {
|
| 372 |
-
background: #161b22 !important;
|
| 373 |
-
border: 1px solid #21262d !important;
|
| 374 |
-
border-radius: 8px !important;
|
| 375 |
-
}
|
| 376 |
-
|
| 377 |
-
/* ── Tabs ── */
|
| 378 |
.tab-nav button {
|
| 379 |
-
font-family: 'Inter', sans-serif !important;
|
| 380 |
font-weight: 500 !important;
|
| 381 |
-
font-size: 0.9rem !important;
|
| 382 |
-
color: #8b949e !important;
|
| 383 |
border: none !important;
|
| 384 |
-
background: transparent !important;
|
| 385 |
-
padding: 10px 20px !important;
|
| 386 |
border-bottom: 2px solid transparent !important;
|
| 387 |
}
|
| 388 |
.tab-nav button.selected {
|
| 389 |
color: #58a6ff !important;
|
| 390 |
-
border-bottom:
|
| 391 |
-
}
|
| 392 |
-
|
| 393 |
-
/* ── Sliders ── */
|
| 394 |
-
input[type="range"] {
|
| 395 |
-
accent-color: #58a6ff !important;
|
| 396 |
-
}
|
| 397 |
-
.gr-slider label {
|
| 398 |
-
font-weight: 500 !important;
|
| 399 |
-
}
|
| 400 |
-
|
| 401 |
-
/* ── Buttons ── */
|
| 402 |
-
.gr-button-primary {
|
| 403 |
-
background: #238636 !important;
|
| 404 |
-
border: 1px solid #2ea043 !important;
|
| 405 |
-
color: #ffffff !important;
|
| 406 |
-
font-family: 'Inter', sans-serif !important;
|
| 407 |
-
font-weight: 500 !important;
|
| 408 |
-
border-radius: 6px !important;
|
| 409 |
-
padding: 8px 24px !important;
|
| 410 |
-
transition: background 0.15s ease !important;
|
| 411 |
-
}
|
| 412 |
-
.gr-button-primary:hover {
|
| 413 |
-
background: #2ea043 !important;
|
| 414 |
-
}
|
| 415 |
-
|
| 416 |
-
/* ── Radio buttons ── */
|
| 417 |
-
.gr-radio label {
|
| 418 |
-
font-family: 'Inter', sans-serif !important;
|
| 419 |
-
}
|
| 420 |
-
|
| 421 |
-
/* ── Audio player ── */
|
| 422 |
-
audio {
|
| 423 |
-
border-radius: 6px !important;
|
| 424 |
-
}
|
| 425 |
-
|
| 426 |
-
/* ── Tables in markdown ── */
|
| 427 |
-
table {
|
| 428 |
-
font-family: 'Inter', sans-serif !important;
|
| 429 |
-
font-size: 0.82rem !important;
|
| 430 |
-
border-collapse: collapse !important;
|
| 431 |
-
width: 100% !important;
|
| 432 |
-
}
|
| 433 |
-
th {
|
| 434 |
-
background: #21262d !important;
|
| 435 |
-
color: #8b949e !important;
|
| 436 |
-
font-weight: 500 !important;
|
| 437 |
-
padding: 8px 12px !important;
|
| 438 |
-
text-align: left !important;
|
| 439 |
-
}
|
| 440 |
-
td {
|
| 441 |
-
padding: 6px 12px !important;
|
| 442 |
-
border-top: 1px solid #21262d !important;
|
| 443 |
-
color: #c9d1d9 !important;
|
| 444 |
-
}
|
| 445 |
-
|
| 446 |
-
/* ── Info text under sliders ── */
|
| 447 |
-
.gr-info {
|
| 448 |
-
font-size: 0.78rem !important;
|
| 449 |
-
color: #6e7681 !important;
|
| 450 |
-
font-style: normal !important;
|
| 451 |
-
}
|
| 452 |
-
|
| 453 |
-
/* ── Code / mono ── */
|
| 454 |
-
code, .mono {
|
| 455 |
-
font-family: 'JetBrains Mono', monospace !important;
|
| 456 |
-
font-size: 0.82rem !important;
|
| 457 |
-
background: #21262d !important;
|
| 458 |
-
padding: 2px 6px !important;
|
| 459 |
-
border-radius: 4px !important;
|
| 460 |
-
}
|
| 461 |
-
|
| 462 |
-
/* ── Accent color for links ── */
|
| 463 |
-
a { color: #58a6ff !important; }
|
| 464 |
-
|
| 465 |
-
/* ── Divider ── */
|
| 466 |
-
hr {
|
| 467 |
-
border: none !important;
|
| 468 |
-
border-top: 1px solid #21262d !important;
|
| 469 |
-
margin: 1.5rem 0 !important;
|
| 470 |
}
|
|
|
|
|
|
|
|
|
|
| 471 |
"""
|
| 472 |
|
| 473 |
-
|
| 474 |
# ---------------------------------------------------------------------------
|
| 475 |
# UI
|
| 476 |
# ---------------------------------------------------------------------------
|
| 477 |
|
| 478 |
-
with gr.Blocks(
|
| 479 |
-
title="Granular Synthesis · Rain",
|
| 480 |
-
css=CUSTOM_CSS,
|
| 481 |
-
theme=gr.themes.Base(),
|
| 482 |
-
) as demo:
|
| 483 |
|
| 484 |
gr.Markdown(
|
| 485 |
"""
|
| 486 |
-
# Granular Synthesis
|
| 487 |
-
###
|
| 488 |
"""
|
| 489 |
)
|
| 490 |
|
| 491 |
with gr.Tabs():
|
| 492 |
|
| 493 |
-
# ======================== RAIN
|
| 494 |
with gr.TabItem("Rain Simulation"):
|
| 495 |
gr.Markdown(
|
| 496 |
"""
|
| 497 |
-
|
| 498 |
-
|
| 499 |
-
|
| 500 |
-
|
| 501 |
-
|
| 502 |
"""
|
| 503 |
)
|
| 504 |
|
|
@@ -508,37 +514,37 @@ with gr.Blocks(
|
|
| 508 |
choices=["light", "medium", "heavy", "thunder"],
|
| 509 |
value="medium",
|
| 510 |
label="Rain type",
|
| 511 |
-
info="
|
| 512 |
-
)
|
| 513 |
-
drop_size = gr.Slider(
|
| 514 |
-
minimum=2, maximum=60, value=18, step=1,
|
| 515 |
-
label="Drop size (ms)",
|
| 516 |
-
info="Grain duration. Larger → splashier, more resonant.",
|
| 517 |
-
)
|
| 518 |
-
rain_density = gr.Slider(
|
| 519 |
-
minimum=10, maximum=400, value=100, step=5,
|
| 520 |
-
label="Drops / second",
|
| 521 |
-
info="Base rate before type multiplier.",
|
| 522 |
-
)
|
| 523 |
-
rain_intensity = gr.Slider(
|
| 524 |
-
minimum=0.5, maximum=5.0, value=1.5, step=0.1,
|
| 525 |
-
label="Intensity",
|
| 526 |
-
info="Global density and wash volume scaling.",
|
| 527 |
)
|
| 528 |
brightness = gr.Slider(
|
| 529 |
-
|
| 530 |
label="Surface brightness",
|
| 531 |
-
info="
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 532 |
)
|
| 533 |
-
|
| 534 |
-
|
| 535 |
-
label="
|
| 536 |
-
info="
|
| 537 |
)
|
| 538 |
stereo = gr.Slider(
|
| 539 |
-
|
| 540 |
label="Stereo width",
|
| 541 |
-
info="0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 542 |
)
|
| 543 |
rain_btn = gr.Button("Generate rain", variant="primary", size="lg")
|
| 544 |
|
|
@@ -547,26 +553,26 @@ with gr.Blocks(
|
|
| 547 |
|
| 548 |
gr.Markdown(
|
| 549 |
"""
|
| 550 |
-
**Presets
|
| 551 |
|
| 552 |
-
| Scene | Type |
|
| 553 |
|---|---|---|---|---|---|---|
|
| 554 |
-
| Drizzle on leaves | light |
|
| 555 |
-
| Window at night | medium |
|
| 556 |
-
| Tin roof | medium |
|
| 557 |
-
| Downpour | heavy |
|
| 558 |
-
| Thunderstorm | thunder |
|
| 559 |
-
| Forest canopy | light |
|
| 560 |
"""
|
| 561 |
)
|
| 562 |
|
| 563 |
rain_btn.click(
|
| 564 |
fn=cb_rain,
|
| 565 |
-
inputs=[rain_type,
|
| 566 |
outputs=rain_audio,
|
| 567 |
)
|
| 568 |
|
| 569 |
-
# ======================== TONAL
|
| 570 |
with gr.TabItem("Tonal Granular"):
|
| 571 |
gr.Markdown(
|
| 572 |
"""
|
|
@@ -578,10 +584,10 @@ with gr.Blocks(
|
|
| 578 |
with gr.Column(scale=1):
|
| 579 |
source_freq = gr.Slider(80, 880, 220, step=1, label="Source frequency (Hz)")
|
| 580 |
grain_size = gr.Slider(5, 200, 50, step=1, label="Grain size (ms)",
|
| 581 |
-
info="Smaller
|
| 582 |
tonal_density = gr.Slider(1, 8, 4, step=0.5, label="Density (overlap)")
|
| 583 |
randomness_sl = gr.Slider(0, 1, 0.3, step=0.01, label="Position randomness",
|
| 584 |
-
info="0
|
| 585 |
pitch = gr.Slider(0.25, 4.0, 1.0, step=0.05, label="Pitch shift")
|
| 586 |
tonal_btn = gr.Button("Synthesize", variant="primary", size="lg")
|
| 587 |
|
|
@@ -591,8 +597,8 @@ with gr.Blocks(
|
|
| 591 |
"""
|
| 592 |
**Signal chain**
|
| 593 |
|
| 594 |
-
Source (additive harmonics)
|
| 595 |
-
|
| 596 |
"""
|
| 597 |
)
|
| 598 |
|
|
@@ -605,7 +611,7 @@ with gr.Blocks(
|
|
| 605 |
gr.Markdown(
|
| 606 |
"""
|
| 607 |
---
|
| 608 |
-
Built with Python and Gradio
|
| 609 |
Part of [Generative Audio Soundscapes Lab](https://my-sonicase.github.io/genaudio-soundscapes/).
|
| 610 |
"""
|
| 611 |
)
|
|
|
|
| 1 |
"""
|
| 2 |
+
Granular Synthesis Demo // Rain Simulation (v3)
|
| 3 |
+
================================================
|
| 4 |
+
|
| 5 |
+
Why v1/v2 sounded like frying bacon:
|
| 6 |
+
The old approach generated individual noise-burst "drops" and summed them.
|
| 7 |
+
This creates a sparse, clicky texture because:
|
| 8 |
+
1. Short noise bursts have flat spectra (white noise = frying sound)
|
| 9 |
+
2. Box/naive filters barely shape the spectrum
|
| 10 |
+
3. Individual grains are too sparse to fuse into a continuous texture
|
| 11 |
+
|
| 12 |
+
Real rain is NOT a sum of isolated clicks. Acoustically, rain is a
|
| 13 |
+
CONTINUOUS stochastic process with a specific spectral shape:
|
| 14 |
+
- Energy concentrated between 1 kHz and 15 kHz
|
| 15 |
+
- Peak around 5-8 kHz (research: Nystuen et al., raindrop acoustics)
|
| 16 |
+
- Spectral slope that varies with rain intensity
|
| 17 |
+
- Slow amplitude modulation (gusts, intensity fluctuation)
|
| 18 |
+
- Small drops produce 13-25 kHz (drizzle shimmer)
|
| 19 |
+
- Large drops add energy below 2 kHz (heavy rain rumble)
|
| 20 |
+
|
| 21 |
+
v3 approach: spectral domain synthesis.
|
| 22 |
+
1. Generate white noise in the frequency domain (FFT)
|
| 23 |
+
2. Sculpt the spectrum to match real rain profiles
|
| 24 |
+
3. Add temporal modulation (amplitude envelopes that breathe)
|
| 25 |
+
4. Layer: continuous wash + transient drops + optional thunder
|
| 26 |
+
5. Stereo decorrelation for spatial width
|
| 27 |
+
|
| 28 |
+
This is still granular thinking: the "grains" are now overlapping
|
| 29 |
+
FFT frames (STFT), each with a shaped spectrum. The overlap-add
|
| 30 |
+
reconstruction is the same principle as classic granular synthesis.
|
| 31 |
"""
|
| 32 |
|
| 33 |
import numpy as np
|
| 34 |
+
from scipy import signal as sig
|
| 35 |
+
from scipy.fft import rfft, irfft
|
| 36 |
import gradio as gr
|
| 37 |
|
| 38 |
# ---------------------------------------------------------------------------
|
| 39 |
# Constants
|
| 40 |
# ---------------------------------------------------------------------------
|
| 41 |
SR = 44100
|
| 42 |
+
DURATION = 7.0 # seconds
|
| 43 |
|
| 44 |
# ---------------------------------------------------------------------------
|
| 45 |
+
# Spectral rain profile
|
| 46 |
# ---------------------------------------------------------------------------
|
| 47 |
|
| 48 |
+
def rain_spectral_profile(
|
| 49 |
+
n_fft: int,
|
| 50 |
+
brightness: float,
|
| 51 |
+
rain_type: str,
|
| 52 |
+
) -> np.ndarray:
|
| 53 |
"""
|
| 54 |
+
Build a frequency-domain magnitude envelope that matches
|
| 55 |
+
the spectral shape of real rainfall.
|
| 56 |
|
| 57 |
+
Based on underwater acoustic rainfall studies:
|
| 58 |
+
small drops peak at 13-25 kHz, large drops are broadband 1-50 kHz,
|
| 59 |
+
most rain energy sits in the 2-12 kHz band.
|
|
|
|
|
|
|
| 60 |
|
| 61 |
+
We model this as a bandpass profile (skewed Gaussian in log-frequency)
|
| 62 |
+
whose center frequency and bandwidth shift with brightness and rain type.
|
| 63 |
"""
|
| 64 |
+
n_bins = n_fft // 2 + 1
|
| 65 |
+
freqs = np.linspace(0, SR / 2, n_bins)
|
| 66 |
+
|
| 67 |
+
# Avoid log(0)
|
| 68 |
+
freqs_safe = np.maximum(freqs, 1.0)
|
| 69 |
+
log_freqs = np.log2(freqs_safe)
|
| 70 |
+
|
| 71 |
+
# Center frequency shifts with brightness
|
| 72 |
+
# Low brightness (dark/soil): center around 2 kHz
|
| 73 |
+
# High brightness (glass/metal): center around 8 kHz
|
| 74 |
+
center_hz = 1500 * (2.0 ** (brightness * 2.5)) # 1.5 kHz to ~8.5 kHz
|
| 75 |
+
center_log = np.log2(center_hz)
|
| 76 |
+
|
| 77 |
+
# Bandwidth in octaves (wider for heavy rain)
|
| 78 |
+
bw_map = {"light": 1.8, "medium": 2.2, "heavy": 3.0, "thunder": 3.5}
|
| 79 |
+
bw = bw_map.get(rain_type, 2.2)
|
| 80 |
+
|
| 81 |
+
# Skewed Gaussian in log-frequency space
|
| 82 |
+
profile = np.exp(-0.5 * ((log_freqs - center_log) / bw) ** 2)
|
| 83 |
+
|
| 84 |
+
# Add high-frequency shimmer for light rain (drizzle peak at 13-25 kHz)
|
| 85 |
+
if rain_type == "light":
|
| 86 |
+
shimmer_center = np.log2(16000)
|
| 87 |
+
shimmer = 0.4 * np.exp(-0.5 * ((log_freqs - shimmer_center) / 0.5) ** 2)
|
| 88 |
+
profile += shimmer
|
| 89 |
+
|
| 90 |
+
# Add sub-bass rumble for heavy/thunder
|
| 91 |
+
if rain_type in ("heavy", "thunder"):
|
| 92 |
+
bass_center = np.log2(300)
|
| 93 |
+
bass = 0.3 * np.exp(-0.5 * ((log_freqs - bass_center) / 1.0) ** 2)
|
| 94 |
+
profile += bass
|
| 95 |
+
|
| 96 |
+
# Roll off everything below 80 Hz (rumble is not rain)
|
| 97 |
+
highpass = 1.0 / (1.0 + (80.0 / freqs_safe) ** 4)
|
| 98 |
+
profile *= highpass
|
| 99 |
|
| 100 |
# Normalize
|
| 101 |
+
profile /= np.max(profile) + 1e-12
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
|
| 103 |
+
return profile
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
|
| 105 |
|
| 106 |
# ---------------------------------------------------------------------------
|
| 107 |
+
# Temporal modulation (rain is not perfectly steady)
|
| 108 |
# ---------------------------------------------------------------------------
|
| 109 |
|
| 110 |
+
def make_modulation(n_samples: int, speed: float = 0.2) -> np.ndarray:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
"""
|
| 112 |
+
Slow amplitude modulation to simulate natural intensity fluctuation.
|
| 113 |
+
Rain intensity varies over seconds (gusts, cloud cells passing).
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
|
| 115 |
+
We sum a few slow random sinusoids to create an organic envelope.
|
|
|
|
|
|
|
|
|
|
| 116 |
"""
|
| 117 |
+
t = np.linspace(0, DURATION, n_samples)
|
| 118 |
+
mod = np.ones(n_samples, dtype=np.float64)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
|
| 120 |
+
# Sum of 4 slow sinusoids with random phases
|
| 121 |
+
for i in range(4):
|
| 122 |
+
freq = speed * (0.5 + i * 0.3) + np.random.uniform(-0.05, 0.05)
|
| 123 |
+
phase = np.random.uniform(0, 2 * np.pi)
|
| 124 |
+
depth = 0.08 + 0.07 * i # increasing modulation depth
|
| 125 |
+
mod += depth * np.sin(2 * np.pi * freq * t + phase)
|
| 126 |
+
|
| 127 |
+
# Keep in a reasonable range
|
| 128 |
+
mod = np.clip(mod, 0.3, 1.5)
|
| 129 |
+
# Smooth with a gentle low-pass to avoid sudden jumps
|
| 130 |
+
window = np.hanning(int(SR * 0.3))
|
| 131 |
+
window /= window.sum()
|
| 132 |
+
mod = np.convolve(mod, window, mode="same")
|
| 133 |
+
|
| 134 |
+
return mod
|
| 135 |
|
| 136 |
|
| 137 |
# ---------------------------------------------------------------------------
|
| 138 |
+
# Core: spectral rain synthesis via overlap-add STFT
|
| 139 |
# ---------------------------------------------------------------------------
|
| 140 |
|
| 141 |
def synthesize_rain(
|
| 142 |
rain_type: str,
|
| 143 |
+
brightness: float,
|
| 144 |
+
density: float,
|
| 145 |
+
modulation_speed: float,
|
|
|
|
|
|
|
| 146 |
stereo_width: float,
|
| 147 |
+
highcut: float,
|
| 148 |
+
lowcut: float,
|
| 149 |
) -> np.ndarray:
|
| 150 |
"""
|
| 151 |
+
Synthesize rain using FFT-based spectral shaping.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
|
| 153 |
+
This is granular synthesis at the frame level:
|
| 154 |
+
each STFT frame is a "grain" whose spectrum is sculpted,
|
| 155 |
+
and the overlap-add reconstruction creates the continuous texture.
|
| 156 |
"""
|
| 157 |
n_out = int(DURATION * SR)
|
| 158 |
+
|
| 159 |
+
# FFT parameters
|
| 160 |
+
# 2048 samples at 44.1kHz = ~46ms frames. This is our "grain size"
|
| 161 |
+
# in the spectral domain. Overlap of 75% ensures smooth transitions.
|
| 162 |
+
n_fft = 2048
|
| 163 |
+
hop = n_fft // 4 # 75% overlap (standard for STFT)
|
| 164 |
+
n_frames = (n_out // hop) + 1
|
| 165 |
+
n_bins = n_fft // 2 + 1
|
| 166 |
+
|
| 167 |
+
# Build the target spectral profile
|
| 168 |
+
profile = rain_spectral_profile(n_fft, brightness, rain_type)
|
| 169 |
+
|
| 170 |
+
# Apply density scaling (affects overall energy)
|
| 171 |
+
profile *= (0.3 + density * 0.7)
|
| 172 |
+
|
| 173 |
+
# Apply frequency range limits from sliders
|
| 174 |
+
freqs = np.linspace(0, SR / 2, n_bins)
|
| 175 |
+
# Gentle roll-off at the edges (not a brick wall, which sounds unnatural)
|
| 176 |
+
low_rolloff = 1.0 / (1.0 + (lowcut / (freqs + 1e-6)) ** 6)
|
| 177 |
+
high_rolloff = 1.0 / (1.0 + (freqs / highcut) ** 6)
|
| 178 |
+
profile *= low_rolloff * high_rolloff
|
| 179 |
+
|
| 180 |
+
# Synthesis window (Hann for overlap-add, same as classic granular)
|
| 181 |
+
window = np.hanning(n_fft)
|
| 182 |
+
|
| 183 |
+
# Two independent channels for stereo
|
| 184 |
+
output_L = np.zeros(n_out + n_fft, dtype=np.float64)
|
| 185 |
+
output_R = np.zeros(n_out + n_fft, dtype=np.float64)
|
| 186 |
+
|
| 187 |
+
for frame in range(n_frames):
|
| 188 |
+
# Generate random phase noise in the frequency domain.
|
| 189 |
+
# This is the key insight: white noise = uniform random phase
|
| 190 |
+
# + flat magnitude. By keeping random phase but imposing our
|
| 191 |
+
# spectral profile as magnitude, we get colored noise that
|
| 192 |
+
# matches the rain spectrum exactly.
|
| 193 |
+
|
| 194 |
+
# Left channel
|
| 195 |
+
phase_L = np.random.uniform(0, 2 * np.pi, n_bins)
|
| 196 |
+
spectrum_L = profile * np.exp(1j * phase_L)
|
| 197 |
+
grain_L = irfft(spectrum_L, n=n_fft).real * window
|
| 198 |
+
|
| 199 |
+
# Right channel: independent phase for stereo decorrelation.
|
| 200 |
+
# stereo_width controls how different L and R are.
|
| 201 |
+
# width=0: identical (mono). width=1: fully independent.
|
| 202 |
+
if stereo_width > 0.01:
|
| 203 |
+
phase_R = phase_L * (1 - stereo_width) + np.random.uniform(0, 2 * np.pi, n_bins) * stereo_width
|
| 204 |
+
spectrum_R = profile * np.exp(1j * phase_R)
|
| 205 |
+
grain_R = irfft(spectrum_R, n=n_fft).real * window
|
| 206 |
+
else:
|
| 207 |
+
grain_R = grain_L.copy()
|
| 208 |
+
|
| 209 |
+
# Place grain in output (overlap-add)
|
| 210 |
+
pos = frame * hop
|
| 211 |
+
if pos + n_fft <= len(output_L):
|
| 212 |
+
output_L[pos:pos + n_fft] += grain_L
|
| 213 |
+
output_R[pos:pos + n_fft] += grain_R
|
| 214 |
+
|
| 215 |
+
# Trim to exact length
|
| 216 |
+
output_L = output_L[:n_out]
|
| 217 |
+
output_R = output_R[:n_out]
|
| 218 |
+
|
| 219 |
+
# Apply temporal modulation
|
| 220 |
+
mod = make_modulation(n_out, speed=modulation_speed)
|
| 221 |
+
output_L *= mod
|
| 222 |
+
output_R *= mod
|
| 223 |
+
|
| 224 |
+
# Add transient drop layer for texture (sparse individual drops on top)
|
| 225 |
+
drop_layer_L, drop_layer_R = make_drop_layer(n_out, rain_type, brightness, density)
|
| 226 |
+
# Drops are much quieter than the continuous layer
|
| 227 |
+
drop_mix = {"light": 0.5, "medium": 0.3, "heavy": 0.15, "thunder": 0.1}
|
| 228 |
+
dmix = drop_mix.get(rain_type, 0.3)
|
| 229 |
+
output_L += drop_layer_L * dmix
|
| 230 |
+
output_R += drop_layer_R * dmix
|
| 231 |
+
|
| 232 |
+
# Thunder
|
| 233 |
+
if rain_type == "thunder":
|
| 234 |
+
th_L, th_R = make_thunder(n_out)
|
| 235 |
+
output_L += th_L
|
| 236 |
+
output_R += th_R
|
| 237 |
+
|
| 238 |
+
# Final normalization
|
| 239 |
+
stereo = np.column_stack([output_L, output_R])
|
| 240 |
+
peak = np.max(np.abs(stereo))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 241 |
if peak > 0:
|
| 242 |
+
stereo *= 0.85 / peak
|
| 243 |
|
| 244 |
+
return stereo
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
# ---------------------------------------------------------------------------
|
| 248 |
+
# Transient drop layer (sparse individual drops for texture)
|
| 249 |
+
# ---------------------------------------------------------------------------
|
| 250 |
+
|
| 251 |
+
def make_drop_layer(
|
| 252 |
+
n_out: int,
|
| 253 |
+
rain_type: str,
|
| 254 |
+
brightness: float,
|
| 255 |
+
density: float,
|
| 256 |
+
) -> tuple:
|
| 257 |
+
"""
|
| 258 |
+
Sparse individual drops layered on top of the continuous wash.
|
| 259 |
+
These provide the "pointillistic" detail that makes rain sound alive.
|
| 260 |
+
Without them, the wash alone sounds like generic colored noise.
|
| 261 |
+
"""
|
| 262 |
+
output_L = np.zeros(n_out, dtype=np.float64)
|
| 263 |
+
output_R = np.zeros(n_out, dtype=np.float64)
|
| 264 |
+
|
| 265 |
+
# Number of audible drops (not all rain drops are individually heard)
|
| 266 |
+
drops_per_sec = {"light": 8, "medium": 20, "heavy": 40, "thunder": 50}
|
| 267 |
+
n_drops = int(drops_per_sec.get(rain_type, 20) * density * DURATION)
|
| 268 |
+
|
| 269 |
+
# Drop duration in samples (10-40ms)
|
| 270 |
+
base_dur = int(SR * 0.02)
|
| 271 |
+
|
| 272 |
+
# Cutoff frequency for drops (matches surface brightness)
|
| 273 |
+
cutoff_base = 1000 * (2.0 ** (brightness * 3.0)) # 1kHz to 8kHz
|
| 274 |
+
|
| 275 |
+
for _ in range(n_drops):
|
| 276 |
+
pos = np.random.randint(0, max(n_out - base_dur * 3, 1))
|
| 277 |
+
|
| 278 |
+
# Each drop varies in duration and brightness
|
| 279 |
+
dur = int(base_dur * np.random.uniform(0.5, 2.0))
|
| 280 |
+
dur = max(dur, 64)
|
| 281 |
+
|
| 282 |
+
# Synthesize drop: filtered noise with sharp exponential decay
|
| 283 |
+
t = np.linspace(0, 1, dur)
|
| 284 |
+
envelope = np.exp(-np.random.uniform(8, 20) * t)
|
| 285 |
+
noise = np.random.randn(dur) * envelope
|
| 286 |
+
|
| 287 |
+
# Bandpass filter each drop using scipy
|
| 288 |
+
# Cutoff varies per drop for realism
|
| 289 |
+
this_cutoff = cutoff_base * np.random.uniform(0.5, 1.5)
|
| 290 |
+
this_cutoff = min(this_cutoff, SR * 0.45)
|
| 291 |
+
low = max(this_cutoff * 0.3, 100)
|
| 292 |
+
|
| 293 |
+
try:
|
| 294 |
+
sos = sig.butter(2, [low, this_cutoff], btype="bandpass", fs=SR, output="sos")
|
| 295 |
+
drop = sig.sosfilt(sos, noise)
|
| 296 |
+
except Exception:
|
| 297 |
+
drop = noise # fallback if filter params are out of range
|
| 298 |
+
|
| 299 |
+
# Random amplitude (distance simulation)
|
| 300 |
+
amp = np.random.uniform(0.1, 1.0) ** 1.3
|
| 301 |
+
|
| 302 |
+
# Stereo position
|
| 303 |
+
pan = np.random.uniform(0, 1)
|
| 304 |
+
L_gain = np.cos(pan * np.pi / 2) * amp
|
| 305 |
+
R_gain = np.sin(pan * np.pi / 2) * amp
|
| 306 |
+
|
| 307 |
+
end = min(pos + dur, n_out)
|
| 308 |
+
seg = drop[:end - pos]
|
| 309 |
+
output_L[pos:end] += seg * L_gain
|
| 310 |
+
output_R[pos:end] += seg * R_gain
|
| 311 |
+
|
| 312 |
+
return output_L, output_R
|
| 313 |
+
|
| 314 |
+
|
| 315 |
+
# ---------------------------------------------------------------------------
|
| 316 |
+
# Thunder
|
| 317 |
+
# ---------------------------------------------------------------------------
|
| 318 |
+
|
| 319 |
+
def make_thunder(n_out: int) -> tuple:
|
| 320 |
+
"""Low-frequency rumble events with slow attack and long tail."""
|
| 321 |
+
L = np.zeros(n_out, dtype=np.float64)
|
| 322 |
+
R = np.zeros(n_out, dtype=np.float64)
|
| 323 |
+
|
| 324 |
+
n_events = np.random.randint(1, 3)
|
| 325 |
+
for _ in range(n_events):
|
| 326 |
+
pos = np.random.randint(0, n_out // 2)
|
| 327 |
+
dur = int(SR * np.random.uniform(2.0, 4.0))
|
| 328 |
+
t = np.linspace(0, 1, dur)
|
| 329 |
+
|
| 330 |
+
# Sum of low frequencies with random phases
|
| 331 |
+
rumble = np.zeros(dur)
|
| 332 |
+
for f in [20, 30, 45, 60, 80, 100]:
|
| 333 |
+
phase = np.random.uniform(0, 2 * np.pi)
|
| 334 |
+
rumble += np.sin(2 * np.pi * f * t + phase) * np.random.uniform(0.3, 1.0)
|
| 335 |
+
|
| 336 |
+
# Envelope: slow build, long decay
|
| 337 |
+
env = np.exp(-1.0 * t) * (1 - np.exp(-6 * t))
|
| 338 |
+
rumble *= env * 0.4
|
| 339 |
+
|
| 340 |
+
end = min(pos + dur, n_out)
|
| 341 |
+
seg = rumble[:end - pos]
|
| 342 |
+
|
| 343 |
+
# Slightly different L/R for width
|
| 344 |
+
L[pos:end] += seg * np.random.uniform(0.7, 1.0)
|
| 345 |
+
R[pos:end] += seg * np.random.uniform(0.7, 1.0)
|
| 346 |
+
|
| 347 |
+
return L, R
|
| 348 |
|
| 349 |
|
| 350 |
# ---------------------------------------------------------------------------
|
| 351 |
+
# Tonal granular engine (unchanged)
|
| 352 |
# ---------------------------------------------------------------------------
|
| 353 |
|
| 354 |
def make_tonal_source(freq: float = 220.0, duration: float = 2.0) -> np.ndarray:
|
| 355 |
t = np.linspace(0, duration, int(SR * duration), endpoint=False)
|
| 356 |
+
s = np.zeros_like(t)
|
| 357 |
for k in range(1, 8):
|
| 358 |
+
s += (1.0 / k) * np.sin(2 * np.pi * freq * k * t)
|
| 359 |
+
s /= np.max(np.abs(s))
|
| 360 |
+
return s
|
| 361 |
|
| 362 |
|
| 363 |
def granular_synthesize(source, grain_size_ms, density, randomness, pitch_shift):
|
| 364 |
grain_samples = max(int((grain_size_ms / 1000.0) * SR), 64)
|
| 365 |
window = np.hanning(grain_samples)
|
| 366 |
hop = max(int(grain_samples / density), 1)
|
|
|
|
| 367 |
n_out = int(DURATION * SR)
|
| 368 |
output = np.zeros(n_out, dtype=np.float64)
|
| 369 |
|
|
|
|
| 384 |
rand_pos = np.random.randint(0, max(src_len - grain_samples, 1))
|
| 385 |
start = int(seq_pos * (1 - randomness) + rand_pos * randomness)
|
| 386 |
start = np.clip(start, 0, src_len - grain_samples)
|
|
|
|
| 387 |
grain = pitched[start: start + grain_samples] * window
|
| 388 |
out_pos = i * hop
|
| 389 |
if out_pos + grain_samples > n_out:
|
|
|
|
| 398 |
|
| 399 |
|
| 400 |
# ---------------------------------------------------------------------------
|
| 401 |
+
# Callbacks
|
| 402 |
# ---------------------------------------------------------------------------
|
| 403 |
|
| 404 |
+
def cb_rain(rain_type, brightness, density, mod_speed, stereo, highcut, lowcut):
|
| 405 |
+
audio = synthesize_rain(rain_type, brightness, density, mod_speed, stereo, highcut, lowcut)
|
| 406 |
return (SR, audio.astype(np.float32))
|
| 407 |
|
| 408 |
|
|
|
|
| 413 |
|
| 414 |
|
| 415 |
# ---------------------------------------------------------------------------
|
| 416 |
+
# Theme + CSS
|
| 417 |
# ---------------------------------------------------------------------------
|
| 418 |
|
| 419 |
+
dark_theme = gr.themes.Base(
|
| 420 |
+
primary_hue=gr.themes.colors.blue,
|
| 421 |
+
secondary_hue=gr.themes.colors.slate,
|
| 422 |
+
neutral_hue=gr.themes.colors.slate,
|
| 423 |
+
font=gr.themes.GoogleFont("Inter"),
|
| 424 |
+
font_mono=gr.themes.GoogleFont("JetBrains Mono"),
|
| 425 |
+
).set(
|
| 426 |
+
body_background_fill="#0d1117",
|
| 427 |
+
body_background_fill_dark="#0d1117",
|
| 428 |
+
body_text_color="#c9d1d9",
|
| 429 |
+
body_text_color_dark="#c9d1d9",
|
| 430 |
+
body_text_color_subdued="#8b949e",
|
| 431 |
+
body_text_color_subdued_dark="#8b949e",
|
| 432 |
+
background_fill_primary="#161b22",
|
| 433 |
+
background_fill_primary_dark="#161b22",
|
| 434 |
+
background_fill_secondary="#0d1117",
|
| 435 |
+
background_fill_secondary_dark="#0d1117",
|
| 436 |
+
block_background_fill="#161b22",
|
| 437 |
+
block_background_fill_dark="#161b22",
|
| 438 |
+
block_border_color="#21262d",
|
| 439 |
+
block_border_color_dark="#21262d",
|
| 440 |
+
block_label_text_color="#8b949e",
|
| 441 |
+
block_label_text_color_dark="#8b949e",
|
| 442 |
+
block_title_text_color="#e6edf3",
|
| 443 |
+
block_title_text_color_dark="#e6edf3",
|
| 444 |
+
border_color_primary="#21262d",
|
| 445 |
+
border_color_primary_dark="#21262d",
|
| 446 |
+
button_primary_background_fill="#238636",
|
| 447 |
+
button_primary_background_fill_dark="#238636",
|
| 448 |
+
button_primary_background_fill_hover="#2ea043",
|
| 449 |
+
button_primary_background_fill_hover_dark="#2ea043",
|
| 450 |
+
button_primary_text_color="#ffffff",
|
| 451 |
+
button_primary_text_color_dark="#ffffff",
|
| 452 |
+
button_secondary_background_fill="#21262d",
|
| 453 |
+
button_secondary_background_fill_dark="#21262d",
|
| 454 |
+
button_secondary_text_color="#c9d1d9",
|
| 455 |
+
button_secondary_text_color_dark="#c9d1d9",
|
| 456 |
+
input_background_fill="#0d1117",
|
| 457 |
+
input_background_fill_dark="#0d1117",
|
| 458 |
+
input_border_color="#30363d",
|
| 459 |
+
input_border_color_dark="#30363d",
|
| 460 |
+
slider_color="#58a6ff",
|
| 461 |
+
slider_color_dark="#58a6ff",
|
| 462 |
+
link_text_color="#58a6ff",
|
| 463 |
+
link_text_color_dark="#58a6ff",
|
| 464 |
+
)
|
| 465 |
|
| 466 |
+
CUSTOM_CSS = """
|
| 467 |
footer { display: none !important; }
|
| 468 |
+
h1 { letter-spacing: -0.03em !important; font-weight: 600 !important; }
|
| 469 |
+
h3 { color: #8b949e !important; font-weight: 400 !important; }
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 470 |
.tab-nav button {
|
|
|
|
| 471 |
font-weight: 500 !important;
|
|
|
|
|
|
|
| 472 |
border: none !important;
|
|
|
|
|
|
|
| 473 |
border-bottom: 2px solid transparent !important;
|
| 474 |
}
|
| 475 |
.tab-nav button.selected {
|
| 476 |
color: #58a6ff !important;
|
| 477 |
+
border-bottom-color: #58a6ff !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 478 |
}
|
| 479 |
+
table { font-size: 0.82rem !important; }
|
| 480 |
+
th { background: #21262d !important; color: #8b949e !important; font-weight: 500 !important; }
|
| 481 |
+
td { border-top: 1px solid #21262d !important; }
|
| 482 |
"""
|
| 483 |
|
|
|
|
| 484 |
# ---------------------------------------------------------------------------
|
| 485 |
# UI
|
| 486 |
# ---------------------------------------------------------------------------
|
| 487 |
|
| 488 |
+
with gr.Blocks(title="Granular Synthesis", css=CUSTOM_CSS, theme=dark_theme) as demo:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 489 |
|
| 490 |
gr.Markdown(
|
| 491 |
"""
|
| 492 |
+
# Granular Synthesis
|
| 493 |
+
### micro-sound, grain clouds, texture design
|
| 494 |
"""
|
| 495 |
)
|
| 496 |
|
| 497 |
with gr.Tabs():
|
| 498 |
|
| 499 |
+
# ======================== RAIN ========================
|
| 500 |
with gr.TabItem("Rain Simulation"):
|
| 501 |
gr.Markdown(
|
| 502 |
"""
|
| 503 |
+
Spectral-domain rain synthesis. Instead of summing noise clicks,
|
| 504 |
+
we sculpt white noise in the frequency domain to match the spectral
|
| 505 |
+
profile of real rainfall (energy concentrated 2 to 12 kHz, slope varies
|
| 506 |
+
with intensity). Each STFT frame is a "grain" whose spectrum is shaped
|
| 507 |
+
by the rain profile, then overlap-added into the output.
|
| 508 |
"""
|
| 509 |
)
|
| 510 |
|
|
|
|
| 514 |
choices=["light", "medium", "heavy", "thunder"],
|
| 515 |
value="medium",
|
| 516 |
label="Rain type",
|
| 517 |
+
info="Controls spectral shape, transient density, and optional layers.",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 518 |
)
|
| 519 |
brightness = gr.Slider(
|
| 520 |
+
0.0, 1.0, 0.45, step=0.01,
|
| 521 |
label="Surface brightness",
|
| 522 |
+
info="0 = dark (earth, foliage). 1 = bright (glass, tin roof).",
|
| 523 |
+
)
|
| 524 |
+
rain_density = gr.Slider(
|
| 525 |
+
0.2, 3.0, 1.0, step=0.1,
|
| 526 |
+
label="Density",
|
| 527 |
+
info="Overall thickness of the rain texture.",
|
| 528 |
)
|
| 529 |
+
mod_speed = gr.Slider(
|
| 530 |
+
0.05, 1.0, 0.2, step=0.05,
|
| 531 |
+
label="Modulation speed",
|
| 532 |
+
info="How fast the rain intensity fluctuates (gusts).",
|
| 533 |
)
|
| 534 |
stereo = gr.Slider(
|
| 535 |
+
0.0, 1.0, 0.7, step=0.01,
|
| 536 |
label="Stereo width",
|
| 537 |
+
info="0 = mono. 1 = fully decorrelated L/R.",
|
| 538 |
+
)
|
| 539 |
+
lowcut = gr.Slider(
|
| 540 |
+
50, 2000, 150, step=10,
|
| 541 |
+
label="Low cut (Hz)",
|
| 542 |
+
info="Remove frequencies below this point.",
|
| 543 |
+
)
|
| 544 |
+
highcut = gr.Slider(
|
| 545 |
+
2000, 20000, 14000, step=100,
|
| 546 |
+
label="High cut (Hz)",
|
| 547 |
+
info="Remove frequencies above this point.",
|
| 548 |
)
|
| 549 |
rain_btn = gr.Button("Generate rain", variant="primary", size="lg")
|
| 550 |
|
|
|
|
| 553 |
|
| 554 |
gr.Markdown(
|
| 555 |
"""
|
| 556 |
+
**Presets**
|
| 557 |
|
| 558 |
+
| Scene | Type | Bright | Dens | Mod | LoCut | HiCut |
|
| 559 |
|---|---|---|---|---|---|---|
|
| 560 |
+
| Drizzle on leaves | light | 0.2 | 0.6 | 0.1 | 200 | 18000 |
|
| 561 |
+
| Window at night | medium | 0.5 | 1.0 | 0.2 | 150 | 14000 |
|
| 562 |
+
| Tin roof | medium | 0.9 | 1.2 | 0.15 | 300 | 16000 |
|
| 563 |
+
| Downpour | heavy | 0.4 | 2.0 | 0.3 | 100 | 12000 |
|
| 564 |
+
| Thunderstorm | thunder | 0.35 | 2.5 | 0.4 | 80 | 10000 |
|
| 565 |
+
| Forest canopy | light | 0.15 | 0.5 | 0.08 | 200 | 15000 |
|
| 566 |
"""
|
| 567 |
)
|
| 568 |
|
| 569 |
rain_btn.click(
|
| 570 |
fn=cb_rain,
|
| 571 |
+
inputs=[rain_type, brightness, rain_density, mod_speed, stereo, highcut, lowcut],
|
| 572 |
outputs=rain_audio,
|
| 573 |
)
|
| 574 |
|
| 575 |
+
# ======================== TONAL ========================
|
| 576 |
with gr.TabItem("Tonal Granular"):
|
| 577 |
gr.Markdown(
|
| 578 |
"""
|
|
|
|
| 584 |
with gr.Column(scale=1):
|
| 585 |
source_freq = gr.Slider(80, 880, 220, step=1, label="Source frequency (Hz)")
|
| 586 |
grain_size = gr.Slider(5, 200, 50, step=1, label="Grain size (ms)",
|
| 587 |
+
info="Smaller = buzzy. Larger = smooth.")
|
| 588 |
tonal_density = gr.Slider(1, 8, 4, step=0.5, label="Density (overlap)")
|
| 589 |
randomness_sl = gr.Slider(0, 1, 0.3, step=0.01, label="Position randomness",
|
| 590 |
+
info="0 = sequential. 1 = fully random (freeze/texture).")
|
| 591 |
pitch = gr.Slider(0.25, 4.0, 1.0, step=0.05, label="Pitch shift")
|
| 592 |
tonal_btn = gr.Button("Synthesize", variant="primary", size="lg")
|
| 593 |
|
|
|
|
| 597 |
"""
|
| 598 |
**Signal chain**
|
| 599 |
|
| 600 |
+
Source (additive harmonics) > grain extraction (sequential + random blend)
|
| 601 |
+
> Hann window (click-free edges) > overlap-add > normalize
|
| 602 |
"""
|
| 603 |
)
|
| 604 |
|
|
|
|
| 611 |
gr.Markdown(
|
| 612 |
"""
|
| 613 |
---
|
| 614 |
+
Built with Python, NumPy, SciPy and Gradio. Everything is synthesized from scratch, no samples.
|
| 615 |
Part of [Generative Audio Soundscapes Lab](https://my-sonicase.github.io/genaudio-soundscapes/).
|
| 616 |
"""
|
| 617 |
)
|
requirements.txt
CHANGED
|
@@ -1,2 +1,3 @@
|
|
| 1 |
numpy
|
|
|
|
| 2 |
gradio
|
|
|
|
| 1 |
numpy
|
| 2 |
+
scipy
|
| 3 |
gradio
|