Spaces:
Sleeping
Sleeping
Add synth_generator.py
Browse files- synth_generator.py +507 -0
synth_generator.py
ADDED
|
@@ -0,0 +1,507 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Synthetic drum song generator with known ground-truth samples.
|
| 3 |
+
|
| 4 |
+
Generates realistic drum patterns by:
|
| 5 |
+
1. Synthesizing individual drum samples (kick, snare, hihat, etc.) with controlled parameters
|
| 6 |
+
2. Placing them in musical patterns with velocity variation, timing humanization, and overlap
|
| 7 |
+
3. Optionally mixing with bass/harmony for realistic Demucs testing
|
| 8 |
+
4. Returning both the mix AND the isolated ground-truth samples + onset map
|
| 9 |
+
|
| 10 |
+
This gives us a perfect evaluation setup: we know exactly which samples are where,
|
| 11 |
+
so we can compare extracted samples against the originals.
|
| 12 |
+
"""
|
| 13 |
+
|
| 14 |
+
import numpy as np
|
| 15 |
+
from scipy.signal import butter, filtfilt, lfilter
|
| 16 |
+
from dataclasses import dataclass, field
|
| 17 |
+
from typing import Optional
|
| 18 |
+
import soundfile as sf
|
| 19 |
+
import json
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
@dataclass
|
| 23 |
+
class GroundTruthSample:
|
| 24 |
+
"""A ground-truth drum sample used to build the synthetic song."""
|
| 25 |
+
name: str # e.g. "kick", "snare"
|
| 26 |
+
audio: np.ndarray # the clean one-shot sample
|
| 27 |
+
sr: int
|
| 28 |
+
frequency_range: tuple # (low_hz, high_hz) primary energy band
|
| 29 |
+
|
| 30 |
+
@property
|
| 31 |
+
def duration(self) -> float:
|
| 32 |
+
return len(self.audio) / self.sr
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
@dataclass
|
| 36 |
+
class PlacedHit:
|
| 37 |
+
"""A single hit placed in the timeline."""
|
| 38 |
+
sample_name: str
|
| 39 |
+
onset_time: float # seconds
|
| 40 |
+
velocity: float # 0-1 amplitude multiplier
|
| 41 |
+
audio: np.ndarray # the actual audio placed (with velocity applied)
|
| 42 |
+
sr: int
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
@dataclass
|
| 46 |
+
class SyntheticSong:
|
| 47 |
+
"""A complete synthetic drum song with ground truth."""
|
| 48 |
+
mix: np.ndarray # full mix audio
|
| 49 |
+
drums_only: np.ndarray # drums-only mix
|
| 50 |
+
sr: int
|
| 51 |
+
bpm: float
|
| 52 |
+
duration: float
|
| 53 |
+
samples: dict # {name: GroundTruthSample}
|
| 54 |
+
hits: list # [PlacedHit, ...]
|
| 55 |
+
per_sample_stems: dict # {name: np.ndarray} isolated stems
|
| 56 |
+
pattern_description: str
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 60 |
+
# Sample synthesis (parametric drum sounds)
|
| 61 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 62 |
+
|
| 63 |
+
def _butter_filter(y, sr, fmin=None, fmax=None, order=4):
|
| 64 |
+
"""Apply butterworth bandpass/lowpass/highpass filter."""
|
| 65 |
+
nyq = sr / 2
|
| 66 |
+
if fmin and fmax:
|
| 67 |
+
b, a = butter(order, [fmin / nyq, fmax / nyq], btype='band')
|
| 68 |
+
elif fmin:
|
| 69 |
+
b, a = butter(order, fmin / nyq, btype='high')
|
| 70 |
+
elif fmax:
|
| 71 |
+
b, a = butter(order, fmax / nyq, btype='low')
|
| 72 |
+
else:
|
| 73 |
+
return y
|
| 74 |
+
return filtfilt(b, a, y)
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
def synthesize_kick(sr: int = 44100, pitch: float = 60.0,
|
| 78 |
+
decay: float = 12.0, punch: float = 150.0,
|
| 79 |
+
duration: float = 0.25, noise_amount: float = 0.05) -> np.ndarray:
|
| 80 |
+
"""Synthesize a kick drum: sine sweep + sub thump + click."""
|
| 81 |
+
t = np.arange(int(sr * duration)) / sr
|
| 82 |
+
# Frequency sweep: punch Hz β pitch Hz
|
| 83 |
+
freq = (punch - pitch) * np.exp(-30 * t) + pitch
|
| 84 |
+
phase = 2 * np.pi * np.cumsum(freq / sr)
|
| 85 |
+
body = np.sin(phase) * np.exp(-decay * t)
|
| 86 |
+
# Sub thump
|
| 87 |
+
sub = 0.4 * np.sin(2 * np.pi * pitch * t) * np.exp(-15 * t)
|
| 88 |
+
# Click transient
|
| 89 |
+
click = noise_amount * np.random.randn(len(t)) * np.exp(-200 * t)
|
| 90 |
+
click = _butter_filter(click, sr, fmax=4000)
|
| 91 |
+
|
| 92 |
+
kick = body + sub + click
|
| 93 |
+
kick = kick / (np.abs(kick).max() + 1e-8) * 0.95
|
| 94 |
+
return kick.astype(np.float32)
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def synthesize_snare(sr: int = 44100, body_freq: float = 200.0,
|
| 98 |
+
noise_decay: float = 12.0, body_decay: float = 20.0,
|
| 99 |
+
duration: float = 0.25, wire_amount: float = 0.6) -> np.ndarray:
|
| 100 |
+
"""Synthesize a snare drum: body tone + noise wires."""
|
| 101 |
+
t = np.arange(int(sr * duration)) / sr
|
| 102 |
+
# Body
|
| 103 |
+
body = np.sin(2 * np.pi * body_freq * t) * np.exp(-body_decay * t) * 0.5
|
| 104 |
+
# Snare wires (filtered noise)
|
| 105 |
+
noise = np.random.randn(len(t)) * np.exp(-noise_decay * t) * wire_amount
|
| 106 |
+
noise = _butter_filter(noise, sr, fmin=1000, fmax=10000)
|
| 107 |
+
# Overtone ring
|
| 108 |
+
ring = 0.15 * np.sin(2 * np.pi * body_freq * 2.7 * t) * np.exp(-25 * t)
|
| 109 |
+
|
| 110 |
+
snare = body + noise + ring
|
| 111 |
+
snare = snare / (np.abs(snare).max() + 1e-8) * 0.95
|
| 112 |
+
return snare.astype(np.float32)
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def synthesize_hihat(sr: int = 44100, is_open: bool = False,
|
| 116 |
+
brightness: float = 8000.0,
|
| 117 |
+
duration: float = None) -> np.ndarray:
|
| 118 |
+
"""Synthesize a hi-hat: filtered noise with metallic overtones."""
|
| 119 |
+
if duration is None:
|
| 120 |
+
duration = 0.4 if is_open else 0.08
|
| 121 |
+
t = np.arange(int(sr * duration)) / sr
|
| 122 |
+
decay = 6.0 if is_open else 40.0
|
| 123 |
+
|
| 124 |
+
noise = np.random.randn(len(t)) * np.exp(-decay * t)
|
| 125 |
+
noise = _butter_filter(noise, sr, fmin=brightness)
|
| 126 |
+
# Metallic overtones
|
| 127 |
+
metal = 0.2 * np.sin(2 * np.pi * 6500 * t) * np.exp(-(decay + 5) * t)
|
| 128 |
+
metal += 0.1 * np.sin(2 * np.pi * 9200 * t) * np.exp(-(decay + 8) * t)
|
| 129 |
+
|
| 130 |
+
hh = noise + metal
|
| 131 |
+
hh = hh / (np.abs(hh).max() + 1e-8) * 0.7
|
| 132 |
+
return hh.astype(np.float32)
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
def synthesize_tom(sr: int = 44100, pitch: float = 120.0,
|
| 136 |
+
decay: float = 10.0, duration: float = 0.3) -> np.ndarray:
|
| 137 |
+
"""Synthesize a tom: pitched body + slight noise."""
|
| 138 |
+
t = np.arange(int(sr * duration)) / sr
|
| 139 |
+
freq = pitch * 1.5 * np.exp(-8 * t) + pitch
|
| 140 |
+
phase = 2 * np.pi * np.cumsum(freq / sr)
|
| 141 |
+
body = np.sin(phase) * np.exp(-decay * t)
|
| 142 |
+
noise = 0.1 * np.random.randn(len(t)) * np.exp(-20 * t)
|
| 143 |
+
noise = _butter_filter(noise, sr, fmin=200, fmax=3000)
|
| 144 |
+
tom = body + noise
|
| 145 |
+
tom = tom / (np.abs(tom).max() + 1e-8) * 0.9
|
| 146 |
+
return tom.astype(np.float32)
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
def synthesize_cymbal(sr: int = 44100, duration: float = 1.5) -> np.ndarray:
|
| 150 |
+
"""Synthesize a crash/ride cymbal: dense metallic noise."""
|
| 151 |
+
t = np.arange(int(sr * duration)) / sr
|
| 152 |
+
noise = np.random.randn(len(t)) * np.exp(-3 * t)
|
| 153 |
+
noise = _butter_filter(noise, sr, fmin=3000)
|
| 154 |
+
# Multiple metallic partials
|
| 155 |
+
partials = sum(
|
| 156 |
+
(0.15 / (i + 1)) * np.sin(2 * np.pi * f * t) * np.exp(-(2 + i) * t)
|
| 157 |
+
for i, f in enumerate([4200, 5800, 7300, 9100, 11500])
|
| 158 |
+
)
|
| 159 |
+
cym = noise + partials
|
| 160 |
+
cym = cym / (np.abs(cym).max() + 1e-8) * 0.6
|
| 161 |
+
return cym.astype(np.float32)
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
def synthesize_bass_note(sr: int = 44100, freq: float = 65.0,
|
| 165 |
+
duration: float = 0.5) -> np.ndarray:
|
| 166 |
+
"""Synthesize a bass note for adding to the mix (tests Demucs separation)."""
|
| 167 |
+
t = np.arange(int(sr * duration)) / sr
|
| 168 |
+
# Sawtooth-ish bass with harmonics
|
| 169 |
+
wave = (np.sin(2 * np.pi * freq * t) +
|
| 170 |
+
0.5 * np.sin(2 * np.pi * freq * 2 * t) +
|
| 171 |
+
0.25 * np.sin(2 * np.pi * freq * 3 * t))
|
| 172 |
+
envelope = np.minimum(t * 50, 1.0) * np.exp(-3 * t) # quick attack, slow decay
|
| 173 |
+
bass = wave * envelope
|
| 174 |
+
bass = _butter_filter(bass, sr, fmax=500)
|
| 175 |
+
bass = bass / (np.abs(bass).max() + 1e-8) * 0.5
|
| 176 |
+
return bass.astype(np.float32)
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 180 |
+
# Sample set creation with controlled variation
|
| 181 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 182 |
+
|
| 183 |
+
def create_sample_set(sr: int = 44100, seed: int = 42,
|
| 184 |
+
variation: str = "medium") -> dict:
|
| 185 |
+
"""Create a set of ground-truth drum samples with parametric variation.
|
| 186 |
+
|
| 187 |
+
Args:
|
| 188 |
+
variation: "none" (identical hits), "low", "medium", "high"
|
| 189 |
+
"""
|
| 190 |
+
rng = np.random.RandomState(seed)
|
| 191 |
+
|
| 192 |
+
# Base parameters with per-variation noise
|
| 193 |
+
var_scale = {"none": 0.0, "low": 0.05, "medium": 0.15, "high": 0.3}[variation]
|
| 194 |
+
|
| 195 |
+
def vary(val, amount=None):
|
| 196 |
+
a = amount if amount is not None else var_scale
|
| 197 |
+
return val * (1.0 + rng.uniform(-a, a))
|
| 198 |
+
|
| 199 |
+
samples = {
|
| 200 |
+
'kick': GroundTruthSample(
|
| 201 |
+
name='kick',
|
| 202 |
+
audio=synthesize_kick(sr, pitch=vary(60), decay=vary(12), punch=vary(150)),
|
| 203 |
+
sr=sr,
|
| 204 |
+
frequency_range=(30, 300),
|
| 205 |
+
),
|
| 206 |
+
'snare': GroundTruthSample(
|
| 207 |
+
name='snare',
|
| 208 |
+
audio=synthesize_snare(sr, body_freq=vary(200), noise_decay=vary(12)),
|
| 209 |
+
sr=sr,
|
| 210 |
+
frequency_range=(100, 8000),
|
| 211 |
+
),
|
| 212 |
+
'hihat_closed': GroundTruthSample(
|
| 213 |
+
name='hihat_closed',
|
| 214 |
+
audio=synthesize_hihat(sr, is_open=False, brightness=vary(8000)),
|
| 215 |
+
sr=sr,
|
| 216 |
+
frequency_range=(3000, 20000),
|
| 217 |
+
),
|
| 218 |
+
'hihat_open': GroundTruthSample(
|
| 219 |
+
name='hihat_open',
|
| 220 |
+
audio=synthesize_hihat(sr, is_open=True, brightness=vary(7000)),
|
| 221 |
+
sr=sr,
|
| 222 |
+
frequency_range=(2000, 20000),
|
| 223 |
+
),
|
| 224 |
+
'tom': GroundTruthSample(
|
| 225 |
+
name='tom',
|
| 226 |
+
audio=synthesize_tom(sr, pitch=vary(120), decay=vary(10)),
|
| 227 |
+
sr=sr,
|
| 228 |
+
frequency_range=(50, 2000),
|
| 229 |
+
),
|
| 230 |
+
'cymbal': GroundTruthSample(
|
| 231 |
+
name='cymbal',
|
| 232 |
+
audio=synthesize_cymbal(sr),
|
| 233 |
+
sr=sr,
|
| 234 |
+
frequency_range=(2000, 20000),
|
| 235 |
+
),
|
| 236 |
+
}
|
| 237 |
+
return samples
|
| 238 |
+
|
| 239 |
+
|
| 240 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 241 |
+
# Pattern generation
|
| 242 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 243 |
+
|
| 244 |
+
def generate_basic_rock(bars: int = 4) -> dict:
|
| 245 |
+
"""Basic rock pattern. Returns {sample_name: [(beat_position, velocity), ...]}"""
|
| 246 |
+
pattern = {
|
| 247 |
+
'kick': [],
|
| 248 |
+
'snare': [],
|
| 249 |
+
'hihat_closed': [],
|
| 250 |
+
'hihat_open': [],
|
| 251 |
+
}
|
| 252 |
+
for bar in range(bars):
|
| 253 |
+
offset = bar * 4 # 4 beats per bar
|
| 254 |
+
# Kick on 1 and 3
|
| 255 |
+
pattern['kick'].extend([(offset + 0, 0.9), (offset + 2, 0.85)])
|
| 256 |
+
# Snare on 2 and 4
|
| 257 |
+
pattern['snare'].extend([(offset + 1, 0.85), (offset + 3, 0.9)])
|
| 258 |
+
# HH on every 8th note
|
| 259 |
+
for eighth in range(8):
|
| 260 |
+
vel = 0.6 if eighth % 2 == 0 else 0.4 # accented downbeats
|
| 261 |
+
pattern['hihat_closed'].append((offset + eighth * 0.5, vel))
|
| 262 |
+
# Open hat on "& of 4"
|
| 263 |
+
pattern['hihat_open'].append((offset + 3.5, 0.55))
|
| 264 |
+
return pattern
|
| 265 |
+
|
| 266 |
+
|
| 267 |
+
def generate_funk_pattern(bars: int = 4) -> dict:
|
| 268 |
+
"""Funky syncopated pattern with ghost notes."""
|
| 269 |
+
pattern = {
|
| 270 |
+
'kick': [],
|
| 271 |
+
'snare': [],
|
| 272 |
+
'hihat_closed': [],
|
| 273 |
+
'hihat_open': [],
|
| 274 |
+
'tom': [],
|
| 275 |
+
}
|
| 276 |
+
for bar in range(bars):
|
| 277 |
+
o = bar * 4
|
| 278 |
+
# Syncopated kick
|
| 279 |
+
pattern['kick'].extend([
|
| 280 |
+
(o + 0, 0.95), (o + 0.75, 0.6), (o + 2, 0.9), (o + 2.5, 0.7)
|
| 281 |
+
])
|
| 282 |
+
# Snare with ghost notes
|
| 283 |
+
pattern['snare'].extend([
|
| 284 |
+
(o + 1, 0.9), (o + 1.75, 0.3), (o + 3, 0.85), (o + 3.25, 0.25)
|
| 285 |
+
])
|
| 286 |
+
# 16th note hats
|
| 287 |
+
for sixteenth in range(16):
|
| 288 |
+
vel = 0.5 + 0.2 * (sixteenth % 4 == 0)
|
| 289 |
+
pattern['hihat_closed'].append((o + sixteenth * 0.25, vel))
|
| 290 |
+
# Tom fill on last bar
|
| 291 |
+
if bar == bars - 1:
|
| 292 |
+
pattern['tom'].extend([
|
| 293 |
+
(o + 3, 0.8), (o + 3.25, 0.75), (o + 3.5, 0.85), (o + 3.75, 0.9)
|
| 294 |
+
])
|
| 295 |
+
return pattern
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
def generate_halftime_pattern(bars: int = 4) -> dict:
|
| 299 |
+
"""Half-time/trap-influenced pattern."""
|
| 300 |
+
pattern = {
|
| 301 |
+
'kick': [],
|
| 302 |
+
'snare': [],
|
| 303 |
+
'hihat_closed': [],
|
| 304 |
+
'cymbal': [],
|
| 305 |
+
}
|
| 306 |
+
for bar in range(bars):
|
| 307 |
+
o = bar * 4
|
| 308 |
+
# Kick on 1
|
| 309 |
+
pattern['kick'].append((o + 0, 0.95))
|
| 310 |
+
# Occasional double kick
|
| 311 |
+
if bar % 2 == 1:
|
| 312 |
+
pattern['kick'].append((o + 0.5, 0.7))
|
| 313 |
+
# Snare on 3 only (half time)
|
| 314 |
+
pattern['snare'].append((o + 2, 0.9))
|
| 315 |
+
# Fast hats
|
| 316 |
+
for sixteenth in range(16):
|
| 317 |
+
vel = 0.3 + 0.15 * (sixteenth % 2 == 0)
|
| 318 |
+
pattern['hihat_closed'].append((o + sixteenth * 0.25, vel))
|
| 319 |
+
# Crash on bar 1
|
| 320 |
+
if bar == 0:
|
| 321 |
+
pattern['cymbal'].append((o + 0, 0.7))
|
| 322 |
+
return pattern
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
PATTERNS = {
|
| 326 |
+
'rock': generate_basic_rock,
|
| 327 |
+
'funk': generate_funk_pattern,
|
| 328 |
+
'halftime': generate_halftime_pattern,
|
| 329 |
+
}
|
| 330 |
+
|
| 331 |
+
|
| 332 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 333 |
+
# Song assembly
|
| 334 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 335 |
+
|
| 336 |
+
def assemble_song(
|
| 337 |
+
samples: dict,
|
| 338 |
+
pattern: dict,
|
| 339 |
+
sr: int = 44100,
|
| 340 |
+
bpm: float = 120.0,
|
| 341 |
+
humanize_timing_ms: float = 5.0,
|
| 342 |
+
humanize_velocity: float = 0.05,
|
| 343 |
+
add_bass: bool = True,
|
| 344 |
+
bass_notes: list = None,
|
| 345 |
+
room_noise_db: float = -60.0,
|
| 346 |
+
seed: int = 42,
|
| 347 |
+
) -> SyntheticSong:
|
| 348 |
+
"""Assemble a complete synthetic song from samples and pattern."""
|
| 349 |
+
rng = np.random.RandomState(seed)
|
| 350 |
+
beat_dur = 60.0 / bpm
|
| 351 |
+
|
| 352 |
+
# Calculate total duration
|
| 353 |
+
all_beats = []
|
| 354 |
+
for name, events in pattern.items():
|
| 355 |
+
if events:
|
| 356 |
+
all_beats.extend([e[0] for e in events])
|
| 357 |
+
max_beat = max(all_beats) if all_beats else 4
|
| 358 |
+
total_dur = (max_beat + 2) * beat_dur # add 2 beats of tail
|
| 359 |
+
total_samples = int(total_dur * sr)
|
| 360 |
+
|
| 361 |
+
# Initialize stems
|
| 362 |
+
drums_mix = np.zeros(total_samples, dtype=np.float64)
|
| 363 |
+
per_sample = {name: np.zeros(total_samples, dtype=np.float64) for name in samples}
|
| 364 |
+
hits = []
|
| 365 |
+
|
| 366 |
+
# Place each hit
|
| 367 |
+
for sample_name, events in pattern.items():
|
| 368 |
+
if sample_name not in samples:
|
| 369 |
+
continue
|
| 370 |
+
sample = samples[sample_name]
|
| 371 |
+
for beat_pos, velocity in events:
|
| 372 |
+
# Humanize timing
|
| 373 |
+
timing_offset = rng.normal(0, humanize_timing_ms / 1000.0)
|
| 374 |
+
onset_time = beat_pos * beat_dur + timing_offset
|
| 375 |
+
onset_time = max(0, onset_time)
|
| 376 |
+
|
| 377 |
+
# Humanize velocity
|
| 378 |
+
vel = velocity * (1.0 + rng.uniform(-humanize_velocity, humanize_velocity))
|
| 379 |
+
vel = np.clip(vel, 0.05, 1.0)
|
| 380 |
+
|
| 381 |
+
# Place in timeline
|
| 382 |
+
start = int(onset_time * sr)
|
| 383 |
+
audio = sample.audio * vel
|
| 384 |
+
end = min(start + len(audio), total_samples)
|
| 385 |
+
actual_len = end - start
|
| 386 |
+
|
| 387 |
+
if actual_len <= 0:
|
| 388 |
+
continue
|
| 389 |
+
|
| 390 |
+
drums_mix[start:end] += audio[:actual_len]
|
| 391 |
+
per_sample[sample_name][start:end] += audio[:actual_len]
|
| 392 |
+
|
| 393 |
+
hits.append(PlacedHit(
|
| 394 |
+
sample_name=sample_name,
|
| 395 |
+
onset_time=onset_time,
|
| 396 |
+
velocity=vel,
|
| 397 |
+
audio=audio[:actual_len],
|
| 398 |
+
sr=sr,
|
| 399 |
+
))
|
| 400 |
+
|
| 401 |
+
# Optional bass line (tests Demucs separation)
|
| 402 |
+
bass_track = np.zeros(total_samples, dtype=np.float64)
|
| 403 |
+
if add_bass:
|
| 404 |
+
if bass_notes is None:
|
| 405 |
+
# Simple root note bass on beat 1 and 3
|
| 406 |
+
bass_notes_list = [(0, 65), (2, 65), (4, 82), (6, 82)]
|
| 407 |
+
# Repeat for all bars
|
| 408 |
+
n_bars = int(max_beat / 4) + 1
|
| 409 |
+
bass_notes = []
|
| 410 |
+
for bar in range(n_bars):
|
| 411 |
+
for beat, freq in bass_notes_list:
|
| 412 |
+
if beat + bar * 4 <= max_beat:
|
| 413 |
+
bass_notes.append((beat + bar * 4, freq))
|
| 414 |
+
|
| 415 |
+
for beat_pos, freq in bass_notes:
|
| 416 |
+
onset = beat_pos * beat_dur
|
| 417 |
+
start = int(onset * sr)
|
| 418 |
+
bass = synthesize_bass_note(sr, freq=freq, duration=beat_dur * 2)
|
| 419 |
+
end = min(start + len(bass), total_samples)
|
| 420 |
+
bass_track[start:end] += bass[:end - start]
|
| 421 |
+
|
| 422 |
+
# Add room noise
|
| 423 |
+
noise = rng.randn(total_samples) * (10 ** (room_noise_db / 20))
|
| 424 |
+
|
| 425 |
+
# Final mix
|
| 426 |
+
full_mix = drums_mix + bass_track + noise
|
| 427 |
+
|
| 428 |
+
# Normalize
|
| 429 |
+
peak = np.abs(full_mix).max()
|
| 430 |
+
if peak > 0:
|
| 431 |
+
scale = 0.9 / peak
|
| 432 |
+
full_mix *= scale
|
| 433 |
+
drums_mix *= scale
|
| 434 |
+
for name in per_sample:
|
| 435 |
+
per_sample[name] *= scale
|
| 436 |
+
|
| 437 |
+
return SyntheticSong(
|
| 438 |
+
mix=full_mix.astype(np.float32),
|
| 439 |
+
drums_only=drums_mix.astype(np.float32),
|
| 440 |
+
sr=sr,
|
| 441 |
+
bpm=bpm,
|
| 442 |
+
duration=total_dur,
|
| 443 |
+
samples=samples,
|
| 444 |
+
hits=hits,
|
| 445 |
+
per_sample_stems=per_sample,
|
| 446 |
+
pattern_description=str({k: len(v) for k, v in pattern.items()}),
|
| 447 |
+
)
|
| 448 |
+
|
| 449 |
+
|
| 450 |
+
def generate_test_song(
|
| 451 |
+
pattern_name: str = 'rock',
|
| 452 |
+
bars: int = 4,
|
| 453 |
+
bpm: float = 120.0,
|
| 454 |
+
sr: int = 44100,
|
| 455 |
+
variation: str = 'medium',
|
| 456 |
+
add_bass: bool = True,
|
| 457 |
+
seed: int = 42,
|
| 458 |
+
) -> SyntheticSong:
|
| 459 |
+
"""High-level function: generate a complete test song with ground truth."""
|
| 460 |
+
samples = create_sample_set(sr=sr, seed=seed, variation=variation)
|
| 461 |
+
pattern_fn = PATTERNS.get(pattern_name, generate_basic_rock)
|
| 462 |
+
pattern = pattern_fn(bars=bars)
|
| 463 |
+
|
| 464 |
+
return assemble_song(
|
| 465 |
+
samples=samples,
|
| 466 |
+
pattern=pattern,
|
| 467 |
+
sr=sr,
|
| 468 |
+
bpm=bpm,
|
| 469 |
+
add_bass=add_bass,
|
| 470 |
+
seed=seed,
|
| 471 |
+
)
|
| 472 |
+
|
| 473 |
+
|
| 474 |
+
def save_ground_truth(song: SyntheticSong, output_dir: str):
|
| 475 |
+
"""Save all ground truth data for evaluation."""
|
| 476 |
+
import os
|
| 477 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 478 |
+
os.makedirs(os.path.join(output_dir, 'gt_samples'), exist_ok=True)
|
| 479 |
+
os.makedirs(os.path.join(output_dir, 'gt_stems'), exist_ok=True)
|
| 480 |
+
|
| 481 |
+
# Save mix and drums
|
| 482 |
+
sf.write(os.path.join(output_dir, 'mix.wav'), song.mix, song.sr, subtype='PCM_24')
|
| 483 |
+
sf.write(os.path.join(output_dir, 'drums_only.wav'), song.drums_only, song.sr, subtype='PCM_24')
|
| 484 |
+
|
| 485 |
+
# Save individual samples
|
| 486 |
+
for name, sample in song.samples.items():
|
| 487 |
+
sf.write(os.path.join(output_dir, 'gt_samples', f'{name}.wav'),
|
| 488 |
+
sample.audio, sample.sr, subtype='PCM_24')
|
| 489 |
+
|
| 490 |
+
# Save per-sample stems
|
| 491 |
+
for name, stem in song.per_sample_stems.items():
|
| 492 |
+
sf.write(os.path.join(output_dir, 'gt_stems', f'{name}_stem.wav'),
|
| 493 |
+
stem, song.sr, subtype='PCM_24')
|
| 494 |
+
|
| 495 |
+
# Save hit map
|
| 496 |
+
hit_map = [
|
| 497 |
+
{'sample': h.sample_name, 'onset': h.onset_time, 'velocity': h.velocity}
|
| 498 |
+
for h in song.hits
|
| 499 |
+
]
|
| 500 |
+
with open(os.path.join(output_dir, 'hit_map.json'), 'w') as f:
|
| 501 |
+
json.dump({
|
| 502 |
+
'bpm': song.bpm,
|
| 503 |
+
'duration': song.duration,
|
| 504 |
+
'sr': song.sr,
|
| 505 |
+
'pattern': song.pattern_description,
|
| 506 |
+
'hits': hit_map,
|
| 507 |
+
}, f, indent=2)
|