Spectral Carving
A script that hides images or watermarks in the spectral data of a song - written in response to idiotinium's plugin-based version, which felt more complicated than it needed to be. This does the same thing with a simple python function call.
Needs numpy, scipy, matplotlib and Pillow (PIL).
import numpy as np
import matplotlib.pyplot as plt
from scipy.io import wavfile
from scipy.signal import stft, istft
from PIL import Image, ImageFilter
def scale_image(img, scaling_type):
h,_ = img.shape
y_indices = np.linspace(1, h, h) # Avoid log(0) issue
if scaling_type == 'log':
log_y_indices = np.log(y_indices + 1)
elif scaling_type == 'MEL':
log_y_indices = 2595*np.log(y_indices/700 + 1)
else:
print('unknown scaling type. please use log or MEL')
log_y_indices = (log_y_indices - log_y_indices.min()) / (log_y_indices.max() - log_y_indices.min())
log_y_indices *= (h - 1)
log_y_indices = log_y_indices.astype(np.int32)
output_img = np.zeros_like(img)
for y in range(h - 1):
y1, y2 = log_y_indices[y], log_y_indices[y + 1]
if y1 != y2:
output_img[y1:y2] = img[y]
else:
output_img[y1] = img[y]
for y in range(1, h - 1): # Fill missing pixels via interpolation
if np.all(output_img[y] == 0):
output_img[y] = (output_img[y - 1] + output_img[y + 1]) // 2
return output_img
def apply_image_mask_stereo(wav_file, img_file, output_file, time_range=(0, 2), freq_range=(1000, 5000), alpha = 1, blur_radius=0, scaling_type = 'linear'):
sample_rate, audio = wavfile.read(wav_file)
stereo = (audio.ndim == 2)
if stereo:
left, right = audio[:, 0], audio[:, 1]
else:
left = right = audio
img = Image.open(img_file).convert('L')
if blur_radius > 0:
img = img.filter(ImageFilter.GaussianBlur(radius=blur_radius))
def process_channel(channel):
f, t, Zxx = stft(channel, fs=sample_rate, nperseg=2048, noverlap=1536)
img_resized = img.resize((np.sum((t >= time_range[0]) & (t <= time_range[1])),
np.sum((f >= freq_range[0]) & (f <= freq_range[1]))))
if scaling_type == 'log':
img_array = np.flipud(scale_image(np.array(img_resized) / 255.0, 'log'))
elif scaling_type == 'MEL':
img_array = np.flipud(scale_image(np.array(img_resized) / 255.0, 'MEL'))
else:
img_array = np.flipud(np.array(img_resized) / 255.0)
time_mask = (t >= time_range[0]) & (t <= time_range[1])
freq_mask = (f >= freq_range[0]) & (f <= freq_range[1])
plt.figure(figsize=(12, 6)) # Show original STFT (only region of interest)
plt.subplot(2, 1, 1)
plt.title("Original STFT (Region of Interest)")
plt.imshow(np.log1p(np.abs(Zxx[np.ix_(freq_mask, time_mask)])), aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
# Apply image mask to the region of interest
Zxx[np.ix_(freq_mask, time_mask)] *= (1-alpha*img_array) #I'm inverting the image here, so white color deletes frequencies, and black keeps them.
#if you want to delete with black and keep with white, use:
# Zxx[np.ix_(freq_mask, time_mask)] *= alpha*img_array
plt.subplot(2, 1, 2)
plt.title("STFT with Image Mask (Region of Interest)")
plt.imshow(np.log1p(np.abs(Zxx[np.ix_(freq_mask, time_mask)])), aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
plt.show()
_, reconstructed_audio = istft(Zxx, fs=sample_rate, nperseg=2048, noverlap=1536)
return reconstructed_audio
left_audio = process_channel(left)
right_audio = process_channel(right)
min_length = min(len(left_audio), len(right_audio))
stereo_audio = np.vstack((left_audio[:min_length], right_audio[:min_length])).T
stereo_audio = np.int16(stereo_audio / np.max(np.abs(stereo_audio)) * 32767)
wavfile.write(output_file, sample_rate, stereo_audio)
print(f"Audio with image mask saved to {output_file}")
def apply_image_mask_hidden_mono(wav_file, img_file, output_file, time_range=(0, 2), freq_range=(1000, 5000), alpha= 1, blur_radius=0,scaling_type = 'linear'):
sample_rate, audio = wavfile.read(wav_file)
stereo = (audio.ndim == 2)
if stereo:
left, right = audio[:, 0], audio[:, 1]
else:
left = right = audio
img = Image.open(img_file).convert('L')
if blur_radius > 0:
img = img.filter(ImageFilter.GaussianBlur(radius=blur_radius))
def process_channel(channel):
f, t, Zxx = stft(channel, fs=sample_rate, nperseg=2048, noverlap=1536)
img_resized = img.resize((np.sum((t >= time_range[0]) & (t <= time_range[1])),
np.sum((f >= freq_range[0]) & (f <= freq_range[1]))))
if scaling_type == 'log':
img_array = np.flipud(scale_image(np.array(img_resized) / 255.0, 'log'))
elif scaling_type == 'MEL':
img_array = np.flipud(scale_image(np.array(img_resized) / 255.0, 'MEL'))
else:
img_array = np.flipud(np.array(img_resized) / 255.0)
time_mask = (t >= time_range[0]) & (t <= time_range[1])
freq_mask = (f >= freq_range[0]) & (f <= freq_range[1])
mask = np.zeros_like(Zxx, dtype=float)
mask[np.ix_(freq_mask, time_mask)] = alpha*img_array
Zxx *= mask
_, reconstructed_audio = istft(Zxx, fs=sample_rate, nperseg=2048, noverlap=1536)
return reconstructed_audio
left__edited = process_channel(left)
right_edited = process_channel(right)
min_length = min(len(left), len(right), len(right_edited), len(left__edited)) #put some buffer before and after the soundfile, so nothing gets cut off
combined_left = left[:min_length] - right_edited[:min_length]
combined_right = right[:min_length] - left__edited[:min_length]
f, t, Zxx_left = stft(combined_left, fs=sample_rate, nperseg=2048, noverlap=1536)
time_mask = (t >= time_range[0]) & (t <= time_range[1])
freq_mask = (f >= freq_range[0]) & (f <= freq_range[1])
plt.figure(figsize=(12, 8))
plt.subplot(3, 1, 1)
plt.title("Left Channel STFT")
plt.imshow(np.log1p(np.abs(Zxx_left[np.ix_(freq_mask, time_mask)])), aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
if scaling_type == 'log' or scaling_type == 'MEL':
plt.gca().set_yscale('log')
f, t, Zxx_right = stft(combined_right, fs=sample_rate, nperseg=2048, noverlap=1536)
plt.subplot(3, 1, 2)
plt.title("Right Channel STFT")
plt.imshow(np.log1p(np.abs(Zxx_right[np.ix_(freq_mask, time_mask)])), aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
if scaling_type == 'log' or scaling_type == 'MEL':
plt.gca().set_yscale('log')
plt.subplot(3, 1, 3)
plt.title("Mono (Combined) STFT")
f, t, Zxx_mono = stft((combined_right+combined_left)/2, fs=sample_rate, nperseg=2048, noverlap=1536)
plt.imshow(np.log1p(np.abs(Zxx_mono[np.ix_(freq_mask, time_mask)])), aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
if scaling_type == 'log' or scaling_type == 'MEL':
plt.gca().set_yscale('log')
plt.tight_layout()
plt.show()
stereo_audio = np.vstack((combined_left, combined_right)).T
stereo_audio = np.int16(stereo_audio / np.max(np.abs(stereo_audio)) * 32767)
wavfile.write(output_file, sample_rate, stereo_audio)
print(f"Audio with image mask saved to {output_file}")
def analyze_file(wav_file, time_range, freq_range, log_scale=False):
sample_rate, audio = wavfile.read(wav_file)
stereo = (audio.ndim == 2)
if stereo:
left, right = audio[:, 0], audio[:, 1]
else:
left = right = audio
# Process left channel
f, t, Zxx_left = stft(left, fs=sample_rate, nperseg=2048, noverlap=1536)
time_mask = (t >= time_range[0]) & (t <= time_range[1])
freq_mask = (f >= freq_range[0]) & (f <= freq_range[1])
plt.figure(figsize=(12, 6))
plt.subplot(3, 1, 1)
plt.title("Left Channel STFT")
plt.imshow(np.log1p(np.abs(Zxx_left[np.ix_(freq_mask, time_mask)])),
aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
if log_scale:
plt.gca().set_yscale('log')
# Process right channel
f, t, Zxx_right = stft(right, fs=sample_rate, nperseg=2048, noverlap=1536)
plt.subplot(3, 1, 2)
plt.title("Right Channel STFT")
plt.imshow(np.log1p(np.abs(Zxx_right[np.ix_(freq_mask, time_mask)])),
aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
if log_scale:
plt.gca().set_yscale('log')
# Process mono channel (combined)
f, t, Zxx_mono = stft((right + left) / 2, fs=sample_rate, nperseg=2048, noverlap=1536)
plt.subplot(3, 1, 3)
plt.title("Mono (Combined) STFT")
plt.imshow(np.log1p(np.abs(Zxx_mono[np.ix_(freq_mask, time_mask)])),
aspect='auto', origin='lower',
extent=[time_range[0], time_range[1], freq_range[0], freq_range[1]], cmap='turbo')
plt.colorbar(label='Log Magnitude')
plt.xlabel('Time [s]')
plt.ylabel('Frequency [Hz]')
if log_scale:
plt.gca().set_yscale('log')
plt.tight_layout()
plt.show()
# the hidden mono function subtracts the image created on the right channel from the let channel, and the left channel image from the righ channel.
# If the data wasnt mono before the changes to the stereo file won't be so easily audible, but also not visible, except if you look at it in mono.
# .png and .jpg are tested and work well. But use black and white images for the best results
#most simple use case:
apply_image_mask_hidden_mono('YourSong.wav', 'YourImage.jpg', 'output_hidden_mono.wav')
analyze_file('output_hidden_mono.wav', time_range=(0, 12), freq_range=(10, 10000))
#more advanced usecase, where blur is added to reduce the noise (but makes the image less sharp) and the frequency is scaled logarithmically
apply_image_mask_hidden_mono('YourSong.wav', 'YourImage.jpg', 'output_hidden_mono_advanced.wav', time_range=(4, 8), freq_range=(1000, 8000), alpha = 0.9, blur_radius=5, scaling_type='log') #time_range in s, frequency in Hz
analyze_file('output_hidden_mono.wav', time_range=(0, 12), freq_range=(10, 10000), log_scale=True)
# the stereo function simply carves the image into the spectrum of each channel, the artifacts are easily audible here.
apply_image_mask_stereo('YourSong.wav', 'YourImage.jpg', 'output_stereo.wav') #time_range in s, frequency in Hz
analyze_file('output_stereo.wav', time_range=(0, 20), freq_range=(20, 10000))
#currently only working with .wav files, more support may ot may not be added in the forseeable or unforseeable future, mainly depending on your interest