code to add babble noise to voxforge data

notebooks/noise_addition/add_noise_to_speech.py

+# bit of a hack right now. should be cleaned up for proper use. Stuff hardcoded for voxforge for now.
+import os
+import sys
+import wave
+import audioop
+import numpy as np
+import logging
+import math
+
+from scipy.io import wavfile
+#from scipy import interpolate
+
+
+#ALL_NOISE_TYPES = ['white','pink','hfchannel','babble','factory1','factory2','buccaneer1','buccaneer2','destroyerengine','destroyerops','f16','leopard','m109','machinegun','volvo']
+#nTypes = {'white','babble'};
+#nTypes = {'pink','factory1'};
+#total_num_noise_types = len(ALL_NOISE_TYPES)
+
+
+def load_resample_noise(srcfile, outrate=16000):
+
+    if not os.path.exists(srcfile):
+        logging.warning('Source file not found: {}'.format(srcfile))
+        return None
+
+    try:
+        s_read = wave.open(srcfile, 'r')
+    except:
+        logging.warning('Failed to open file: {}'.format(srcfile))
+        return None
+
+    insamprate =  s_read.getframerate()
+    sampwidth = s_read.getsampwidth()
+    inchannels = s_read.getnchannels()
+    nsamples = s_read.getnframes()
+    indata = s_read.readframes(nsamples)
+
+    s_read.close()
+
+    converted=None
+    if insamprate != outrate:
+        converted, newstate = audioop.ratecv(indata, sampwidth, inchannels, insamprate, outrate, None)
+    else:
+        converted = indata
+        newstate = None
+
+    out_data = np.fromstring(converted, dtype=np.int16)
+    #print('converted:', converted.shape)
+    print(out_data.shape)
+    return out_data
+
+
+def add_babble_noise(audio_sample_file, noise_data, noise_level=10):
+
+    speech_fs, speech_data = wavfile.read(audio_sample_file)
+
+    len_speech = len(speech_data)
+    len_noise = len(noise_data)
+    assert len_noise >= len_speech, 'Noise sample is shorter than speech sample. Quitting'
+    nseg = noise_data[0:len_speech]
+    #nseg = nseg - np.mean(nseg) # It appears that the mean is around 2 % of the max.
+    Pnseg = float(sum(nseg*nseg))/float(len(nseg)) # noise-power
+    my_eps = np.finfo(float).eps
+
+    # Compute the power in dB.    
+    Px = sum(speech_data.astype(float)*speech_data.astype(float))/float(len(speech_data))
+    if(Px < 0):
+        print(Pnseg, Px)
+        print(sum(speech_data*speech_data))
+        print(len(speech_data))
+                
+    factor = math.sqrt(10**(-noise_level/10)*Pnseg/Px)
+    #print('Factor:',factor)
+    noisy_speech = speech_data+(factor*nseg).astype(noise_data.dtype)
+    # Fit it between -1:1. ## don't do it in python. this was good for matlab.
+    #xn = np.round(0.999*xn/(max(abs(xn)) + my_eps)).astype(speech_data.dtype)
+    
+    return speech_fs, noisy_speech
+
+
+def load_speech_file_list():
+    file_list = 'voxforge_denoised_filelist.txt'
+    with open(file_list) as f:
+        content = f.readlines() 
+    f.close()
+    list_of_files = [x.strip() for x in content] 
+    print('no. of speech files:', len(list_of_files))
+
+    return list_of_files 
+
+
+
+
+def main():
+    #data_root = './data'
+    #noise_file = 'conversationNoise_SIGNAL019-20kHz.wav'
+    #speech_file = 'voxforge_samples/enroll/JayCutlersBrother-20080919-ick-a0040-16k.wav'
+    
+    voxforge_root = '/idiap/resource/database/VoxForge/audio/denoised/SpeechCorpus/Trunk/Audio/Main/16kHz_16bit'
+    out_root = '/idiap/temp/sbhatta/work/voxforge_babble'
+    list_of_files = load_speech_file_list()
+#    for i in range(5):
+#        print(list_of_files[i])
+#        my_file = list_of_files[i]
+#        parts = my_file.split(os.sep)
+#        print(parts[1:])
+#        #['.', '1028-20100710-hne', 'wav', 'ar-01.wav']
+
+
+    
+    noise_filename = 'babble_SIGNAL019-20kHz.wav' #os.path.join(data_root, noise_file)
+    noise_data = load_resample_noise(noise_filename, outrate=16000)
+    print('noise loaded')
+    
+    #wavfile.write("./data/test_resampled_noise.wav", speech_fs, noise_data)
+
+    noise_db = 10
+    short_list = list_of_files[0:5]
+    #for i, fv in enumerate(short_list):
+    for i, fv in enumerate(list_of_files):
+        f = os.path.join(voxforge_root, fv)
+
+        parts = fv.split(os.sep)
+        #print(parts[1:])
+        first_dir = os.path.join(out_root, parts[1]) 
+        if not os.path.exists(first_dir):
+            os.makedirs(first_dir)
+            print(first_dir)
+        second_dir = os.path.join(first_dir, parts[2])
+        if not os.path.exists(second_dir):
+            os.makedirs(second_dir)
+        out_f = os.path.join(out_root, fv)
+        if not os.path.isfile(out_f):
+            speech_fs, noisy_speech = add_babble_noise(f, noise_data, noise_level=noise_db)
+            wavfile.write(out_f, speech_fs, noisy_speech)
+
+    #speech_filename = os.path.join(data_root, speech_file)
+    #speech_fs, noisy_speech = add_babble_noise(speech_filename, noise_data, noise_level=10)
+    #wavfile.write("./data/test_resampled_noise.wav", 16000, noisy_speech)
+
+
+if __name__ == "__main__":
+    main()
+
+