Spaces:
Sleeping
Sleeping
File size: 8,511 Bytes
2f54ec8 72a8e1c 2f54ec8 72a8e1c 2f54ec8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 |
import numpy as np
import librosa
import torch
import torch.nn as nn
# import pywt
from scipy import signal
def compute_cwt_power_spectrum(audio, sample_rate, num_freqs=128, f_min=20, f_max=None):
"""
Compute the power spectrum of continuous wavelet transform using Morlet wavelet.
Parameters:
audio: torch.Tensor
Input audio signal
sample_rate: int
Sampling rate of the audio
num_freqs: int
Number of frequency bins for the CWT
f_min: float
Minimum frequency to analyze
f_max: float or None
Maximum frequency to analyze (defaults to Nyquist frequency)
Returns:
torch.Tensor: CWT power spectrum
"""
# Convert to numpy
audio_np = audio.cpu().numpy()
# Set default f_max to Nyquist frequency if not specified
if f_max is None:
f_max = sample_rate // 2
# Generate frequency bins (logarithmically spaced)
frequencies = np.logspace(
np.log10(f_min),
np.log10(f_max),
num=num_freqs
)
# Compute the width of the wavelet (in samples)
widths = sample_rate / (2 * frequencies * np.pi)
# Compute CWT using Morlet wavelet
cwt = signal.cwt(
audio_np,
signal.morlet2,
widths,
w=5.0 # Width parameter of Morlet wavelet
)
# Compute power spectrum (magnitude squared)
power_spectrum = np.abs(cwt) ** 2
# Convert to torch tensor
power_spectrum_tensor = torch.FloatTensor(power_spectrum)
return power_spectrum_tensor
# def compute_wavelet_transform(audio, wavelet, decompos_level):
# """Compute wavelet decomposition of the audio signal."""
# # Convert to numpy and ensure 1D
# audio_np = audio.cpu().numpy()
#
# # Perform wavelet decomposition
# coeffs = pywt.wavedec(audio_np, wavelet, level=decompos_level)
#
# # Stack coefficients into a 2D array
# # First, pad all coefficient arrays to the same length
# max_len = max(len(c) for c in coeffs)
# padded_coeffs = []
# for coeff in coeffs:
# pad_len = max_len - len(coeff)
# if pad_len > 0:
# padded_coeff = np.pad(coeff, (0, pad_len), mode='constant')
# else:
# padded_coeff = coeff
# padded_coeffs.append(padded_coeff)
#
# # Stack into 2D array where each row is a different scale
# wavelet_features = np.stack(padded_coeffs)
#
# # Convert to tensor
# return torch.FloatTensor(wavelet_features)
def compute_melspectrogram(audio, sample_rate):
mel_spec = librosa.feature.melspectrogram(
y=audio.cpu().numpy(),
sr=sample_rate,
n_mels=128
)
return torch.FloatTensor(librosa.power_to_db(mel_spec))
def compute_mfcc(audio, sample_rate):
mfcc = librosa.feature.mfcc(
y=audio.cpu().numpy(),
sr=sample_rate,
n_mfcc=20
)
return torch.FloatTensor(mfcc)
def compute_chroma(audio, sample_rate):
chroma = librosa.feature.chroma_stft(
y=audio.cpu().numpy(),
sr=sample_rate
)
return torch.FloatTensor(chroma)
def compute_time_domain_features(audio, sample_rate, frame_length=2048, hop_length=128):
"""
Compute time-domain features from audio signal.
Returns a dictionary of features.
"""
# Convert to numpy
audio_np = audio.cpu().numpy()
# Initialize dictionary for features
features = {}
# 1. Zero Crossing Rate
zcr = librosa.feature.zero_crossing_rate(
y=audio_np,
frame_length=frame_length,
hop_length=hop_length
)
features['zcr'] = torch.Tensor([zcr.sum()])
# 2. Root Mean Square Energy
rms = librosa.feature.rms(
y=audio_np,
frame_length=frame_length,
hop_length=hop_length
)
features['rms_energy'] = torch.Tensor([rms.mean()])
# 3. Temporal Statistics
frames = librosa.util.frame(audio_np, frame_length=frame_length, hop_length=hop_length)
features['mean'] = torch.Tensor([np.mean(frames, axis=0).mean()])
features['std'] = torch.Tensor([np.std(frames, axis=0).mean()])
features['max'] = torch.Tensor([np.max(frames, axis=0).mean()])
# 4. Tempo and Beat Features
onset_env = librosa.onset.onset_strength(y=audio_np, sr=sample_rate)
tempo = librosa.beat.tempo(onset_envelope=onset_env, sr=sample_rate)
features['tempo'] = torch.Tensor(tempo)
# 5. Amplitude Envelope
envelope = np.abs(librosa.stft(audio_np, n_fft=frame_length, hop_length=hop_length))
features['envelope'] = torch.Tensor([np.mean(envelope, axis=0).mean()])
return features
def compute_frequency_domain_features(audio, sample_rate, n_fft=2048, hop_length=512):
"""
Compute frequency-domain features from audio signal.
Returns a dictionary of features.
"""
# Convert to numpy
audio_np = audio.cpu().numpy()
# Initialize dictionary for features
features = {}
# 1. Spectral Centroid
try:
spectral_centroids = librosa.feature.spectral_centroid(
y=audio_np,
sr=sample_rate,
n_fft=n_fft,
hop_length=hop_length,
)
features['spectral_centroid'] = torch.FloatTensor([spectral_centroids.max()])
except Exception as e:
features['spectral_centroid'] = torch.FloatTensor([np.nan])
# 2. Spectral Rolloff
try:
spectral_rolloff = librosa.feature.spectral_rolloff(
y=audio_np,
sr=sample_rate,
n_fft=n_fft,
hop_length=hop_length,
)
features['spectral_rolloff'] = torch.FloatTensor([spectral_rolloff.max()])
except Exception as e:
features['spectral_rolloff'] = torch.FloatTensor([np.nan])
# 3. Spectral Bandwidth
try:
spectral_bandwidth = librosa.feature.spectral_bandwidth(
y=audio_np,
sr=sample_rate,
n_fft=n_fft,
hop_length=hop_length
)
features['spectral_bandwidth'] = torch.FloatTensor([spectral_bandwidth.max()])
except Exception as e:
features['spectral_bandwidth'] = torch.FloatTensor([np.nan])
# 4. Spectral Contrast
try:
spectral_contrast = librosa.feature.spectral_contrast(
y=audio_np,
sr=sample_rate,
n_fft=n_fft,
hop_length=hop_length,
fmin=20, # Lower minimum frequency
n_bands=4, # Reduce number of bands
quantile=0.02
)
features['spectral_contrast'] = torch.FloatTensor([spectral_contrast.mean()])
except Exception as e:
features['spectral_contrast'] = torch.FloatTensor([np.nan])
# 5. Spectral Flatness
try:
spectral_flatness = librosa.feature.spectral_flatness(
y=audio_np,
n_fft=n_fft,
hop_length=hop_length
)
features['spectral_flatness'] = torch.FloatTensor([spectral_flatness.max()])
except Exception as e:
features['spectral_flatness'] = torch.FloatTensor([np.nan])
# 6. Spectral Flux
try:
stft = np.abs(librosa.stft(audio_np, n_fft=n_fft, hop_length=hop_length))
spectral_flux = np.diff(stft, axis=1)
spectral_flux = np.pad(spectral_flux, ((0, 0), (1, 0)), mode='constant')
features['spectral_flux'] = torch.FloatTensor([np.std(spectral_flux)])
except Exception as e:
features['spectral_flux'] = torch.FloatTensor([np.nan])
return features
def compute_all_features(audio, sample_rate, wavelet='db1', decompos_level=4):
"""
Compute all available features and return them in a dictionary.
"""
features = {}
# Basic transformations
# features['wavelet'] = compute_wavelet_transform(audio, wavelet, decompos_level)
# features['melspectrogram'] = compute_melspectrogram(audio, sample_rate)
# features['mfcc'] = compute_mfcc(audio, sample_rate)
# features['chroma'] = compute_chroma(audio, sample_rate)
# features['cwt_power'] = compute_cwt_power_spectrum(
# audio,
# sample_rate,
# num_freqs=128, # Same as mel bands for consistency
# f_min=20, # Standard lower frequency bound
# f_max=sample_rate // 2 # Nyquist frequency
# )
# Time domain features
# features['time_domain'] = compute_time_domain_features(audio, sample_rate)
# Frequency domain features
features['frequency_domain'] = compute_frequency_domain_features(audio, sample_rate)
return features |