adding source code

UBM.py

@@ -0,0 +1,118 @@
+import os
+import subprocess
+import sidekit
+import numpy as np
+from tqdm import tqdm
+from utils import *
+import multiprocessing as mp
+import warnings
+warnings.filterwarnings("ignore")
+
+
+
+class SpeakerRecognizer():
+
+    def __init__(self):
+        ############ Global Variables ###########
+        self.SAMPLE_RATE = 44100
+        self.NUM_CHANNELS = 2
+        self.PRECISION = 16 #I mean 16-bit
+        self.NUM_THREADS = mp.cpu_count() #(4 default)
+        # Number of Guassian Distributions
+        self.NUM_GUASSIANS = 32
+        ############ Dirs #########
+        self.base_dir = "/media/anwar/E/Voice_Biometrics/SIDEKIT-1.3/py3env"
+        self.input_dir = os.path.join(self.base_dir, "data")
+        self.all_files = os.listdir(self.input_dir)
+        self.wav_dir = os.path.join(self.base_dir, "wav")
+        self.feat_dir = os.path.join(self.base_dir, "feat")
+        self.ubm_dir = os.path.join(self.base_dir, "ubm")
+        self.test_dir = os.path.join(self.base_dir, "test")
+
+
+
+
+    def extractFeatures(self):
+        if not os.path.exists(self.wav_dir):
+            raise Exception("Couldn't extract features, apply preprocess method first!!") 
+
+        # Feature extraction
+        extractor = sidekit.FeaturesExtractor(audio_filename_structure=os.path.join(self.wav_dir, "{}"),
+                                              feature_filename_structure=os.path.join(self.feat_dir, "{}.h5"),
+                                              sampling_frequency=self.SAMPLE_RATE,
+                                              lower_frequency=200,
+                                              higher_frequency=3800,
+                                              filter_bank="log",
+                                              filter_bank_size=24,
+                                              window_size=0.04,
+                                              shift=0.01,
+                                              ceps_number=20,
+                                              vad="snr",
+                                              snr=40,
+                                              pre_emphasis=0.97,
+                                              save_param=["vad", "energy", "cep", "fb"],
+                                              keep_all_features=True)
+
+        # Prepare file lists
+        show_list = np.unique(np.hstack([self.all_files]))
+        channel_list = np.zeros_like(show_list, dtype = int)
+
+        # save the features in feat_dir
+        extractor.save_list(show_list=show_list,
+                            channel_list=channel_list,
+                            num_thread=self.NUM_THREADS)
+
+
+
+    def train(self, SAVE_FLAG=True):
+        #Universal Background Model Training
+        #NOTE: https://projets-lium.univ-lemans.fr/sidekit/tutorial/ubmTraining.html
+        ubm_list = os.listdir(os.path.join(self.base_dir, "feat"))
+        for i in range(len(ubm_list)):
+            ubm_list[i] = ubm_list[i].split(".h5")[0]
+
+        server = sidekit.FeaturesServer(feature_filename_structure=os.path.join(self.feat_dir, "{}.h5"),
+                                        sources=None,
+                                        dataset_list=["vad", "energy", "cep", "fb"],
+                                        feat_norm="cmvn",
+                                        global_cmvn=None,
+                                        dct_pca=False,
+                                        dct_pca_config=None,
+                                        sdc=False,
+                                        sdc_config=None,
+                                        delta=True,
+                                        double_delta=True,
+                                        delta_filter=None,
+                                        context=None,
+                                        traps_dct_nb=None,
+                                        rasta=True,
+                                        keep_all_features=True)
+
+        print("Training...")
+        ubm = sidekit.Mixture()
+        ubm.EM_split(features_server=server,
+                     feature_list=ubm_list,
+                     distrib_nb=self.NUM_GUASSIANS,
+                     iterations=(1, 2, 2, 4, 4, 4, 4, 8, 8, 8, 8, 8, 8),
+                     num_thread=self.NUM_THREADS,
+                     save_partial=True,
+                     ceil_cov=10,
+                     floor_cov=1e-2
+                     )
+        if SAVE_FLAG:
+            modelname = "ubm_{}.h5".format(self.NUM_GUASSIANS)
+            print("Saving the model {} at {}".format(modelname, self.ubm_dir))
+            ubm.write(os.path.join(self.ubm_dir, modelname))
+
+
+    def evaluate(self):
+        pass
+
+
+
+
+if __name__ == "__main__":
+    ubm = SpeakerRecognizer()
+    ubm.preprocess()
+    ubm.extractFeatures()
+    ubm.train()

requirements.txt

@@ -0,0 +1,13 @@
+# requirements file
+
+### BEGIN ###
+matplotlib>=3.0.0
+numpy>=1.15.2
+pyparsing>=2.2.2
+scipy>=1.1.0
+six==1.11.0
+h5py>=2.8.0
+pandas>=0.23.4
+tqdm>=4.31.1
+### END ###
+

utils.py

@@ -0,0 +1,20 @@
+import os
+import subprocess
+
+
+
+def preprocessAudioFile(inwav, outwav, sample_rate, n_channels, bit=16):
+    """
+    This function is used to preprocess audio for SideKit.
+    As we can see, the only default value set is for precision (16-bit)
+    and that's because SideKit has problems for other precisions.
+    And all audio files need to be the same criteria.
+    NOTE:
+    wave.py only reads wav files in PCM format, so make sure that's the case!!
+    """
+    parent, _ = os.path.split(outwav)
+    if not os.path.exists(parent):
+        os.mkdir(parent)
+    command = "sox {} -r {} -c {} -b {} {}"\
+                    .format(inwav, sample_rate, n_channels, bit, outwav)
+    subprocess.call(command, shell=True)