generate_spectrograms.py

import os

import matplotlib
matplotlib.use('agg')

from matplotlib import pyplot as plt
from matplotlib import cm
from tqdm import tqdm
import pylab

import librosa
from librosa import display
import numpy as np

WAV_DIR = 'wav_files/'
IMG_DIR = 'spectrogram_images/'
wav_files = os.listdir(WAV_DIR)

for f in tqdm(wav_files):
    try:
        # Read wav-file
        y, sr = librosa.load(WAV_DIR+f, sr = 22050) # Use the default sampling rate of 22,050 Hz
        
        # Pre-emphasis filter
        pre_emphasis = 0.97
        y = np.append(y[0], y[1:] - pre_emphasis * y[:-1])
        
        # Compute spectrogram
        M = librosa.feature.melspectrogram(y, sr, 
                                           fmax = sr/2, # Maximum frequency to be used on the on the MEL scale
                                           n_fft=2048, 
                                           hop_length=512, 
                                           n_mels = 96, # As per the Google Large-scale audio CNN paper
                                           power = 2) # Power = 2 refers to squared amplitude
        
        # Power in DB
        log_power = librosa.power_to_db(M, ref=np.max)# Covert to dB (log) scale
        
        # Plotting the spectrogram and save as JPG without axes (just the image)
        pylab.figure(figsize=(3,3))
        pylab.axis('off') 
        pylab.axes([0., 0., 1., 1.], frameon=False, xticks=[], yticks=[]) # Remove the white edge
        librosa.display.specshow(log_power, cmap=cm.jet)
        pylab.savefig(IMG_DIR + f[:-4]+'.jpg', bbox_inches=None, pad_inches=0)
        pylab.close()

    except Exception as e:
        print(f, e)
        pass