So I'm wanting to feed spectrogram images to a convolutional neural network as an attempt to classify various sounds. I want each image to be exactly 384x128 pixels. However, when I actually save the image it is only 297x98. Here's my code:
def save_spectrogram(num):
dpi = 128
x_pixels = 384
y_pixels = 128
samples, sr = load_wave(num)
stft = np.absolute(librosa.stft(samples))
db = librosa.amplitude_to_db(stft, ref=np.max)
fig = plt.figure(figsize=(x_pixels//dpi, y_pixels//dpi), dpi=dpi, frameon=False)
ax = fig.add_subplot(111)
ax.axes.get_xaxis().set_visible(False)
ax.axes.get_yaxis().set_visible(False)
ax.set_frame_on(False)
librosa.display.specshow(db, y_axis='linear')
plt.savefig(TRAIN_IMG+str(num)+'.jpg', bbox_inches='tight', pad_inches=0, dpi=dpi)
Does anyone have any pointers on how I can fix this? I've also tried doing it without the subplot, but when I do that it still saves as the wrong size AND has white space/background.
Plots are for humans to look at, and contains things like axis markers, labels etc that are not useful for machine learning. To feed a model with an 'image' of the spectrogram, one should output only the data. This data be stored in any format, but if you want to use a standard image format then should use PNG. Lossy compression such as JPEG introduces compression artifacts.
Here follows working example code to save spectrogram. Note that to get a fixed size image output, the code extracts a fixed-length window of the audio signal. Dividing an audio stream into such fixed-length analysis windows is standard practice.
import librosa
import numpy
import skimage.io
def scale_minmax(X, min=0.0, max=1.0):
X_std = (X - X.min()) / (X.max() - X.min())
X_scaled = X_std * (max - min) + min
return X_scaled
def spectrogram_image(y, sr, out, hop_length, n_mels):
# use log-melspectrogram
mels = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=n_mels,
n_fft=hop_length*2, hop_length=hop_length)
mels = numpy.log(mels + 1e-9) # add small number to avoid log(0)
# min-max scale to fit inside 8-bit range
img = scale_minmax(mels, 0, 255).astype(numpy.uint8)
img = numpy.flip(img, axis=0) # put low frequencies at the bottom in image
img = 255-img # invert. make black==more energy
# save as PNG
skimage.io.imsave(out, img)
if __name__ == '__main__':
# settings
hop_length = 512 # number of samples per time-step in spectrogram
n_mels = 128 # number of bins in spectrogram. Height of image
time_steps = 384 # number of time-steps. Width of image
# load audio. Using example from librosa
path = librosa.util.example_audio_file()
y, sr = librosa.load(path, offset=1.0, duration=10.0, sr=22050)
out = 'out.png'
# extract a fixed length window
start_sample = 0 # starting at beginning
length_samples = time_steps*hop_length
window = y[start_sample:start_sample+length_samples]
# convert to PNG
spectrogram_image(window, sr=sr, out=out, hop_length=hop_length, n_mels=n_mels)
print('wrote file', out)