normalized volume

hparams/hparams.yaml

@@ -1,6 +1,15 @@
 ---
 default_iu3:
+  train2:
+      data_path: 'IU/*_split/*.wav'
   convert:
-      data_path: 'korean/*.wav'
-      one_full_wav: True
-      batch_size: 10
+      one_full_wav: False
+      batch_size: 4
+---
+iu:
+  train2:
+      data_path: 'IU/*_split/*.wav'
+      batch_size: 64
+  convert:
+      one_full_wav: False
+      batch_size: 4

tools/convert_amp.py

@@ -0,0 +1,25 @@
+# -*- coding: utf-8 -*-
+#!/usr/bin/env python
+
+from pydub import AudioSegment
+import glob
+from audio_utils import split_path
+import os
+
+src_path = '/Users/avin/git/vc/datasets/IU/v_app_split'
+target_path = '{}_amp'.format(src_path)
+target_amp_in_db = -20
+
+
+def match_target_amplitude(sound, target_dBFS):
+    change_in_dBFS = target_dBFS - sound.dBFS
+    return sound.apply_gain(change_in_dBFS)
+
+if not os.path.exists(target_path):
+    os.mkdir(target_path)
+
+for filepath in glob.glob('{}/*.wav'.format(src_path)):
+    basepath, filename, _ = split_path(filepath)
+    sound = AudioSegment.from_wav(filepath)
+    normalized_sound = match_target_amplitude(sound, target_amp_in_db)
+    normalized_sound.export('{}/{}.wav'.format(target_path, filename), 'wav')

tools/mfcc_amplitude_normalization_test.py → tools/mfcc_amp_normalization_test.py

@@ -21,7 +21,7 @@
                  n_fft=n_fft,
                  hop_length=hop_length,
                  win_length=win_length)
-mag = np.abs(D)
+mag = np.abs(D)  # (1+n_fft//2, t)
 scaled_mag = mag * 200
 
 # Get mel-spectrogram

tools/plot_spectrogram.py

@@ -7,11 +7,11 @@
 import matplotlib.pyplot as plt
 from audio_utils import read, write
 
-filename = '/Users/avin/git/vc/datasets/timit/TIMIT/TEST/DR1/FAKS0/SA1.wav'
+filename = '/Users/avin/git/vc/outputs/male.wav'
 sr = 22050
-n_fft = 4096
+n_fft = 1024
 len_hop = n_fft / 4
-plot_wav = True
+plot_wav = False
 plot_spec = True
 
 # Waveforms
@@ -37,7 +37,7 @@
 if plot_spec:
     plt.figure(2)
 
-    librosa.display.specshow(librosa.amplitude_to_db(spec, ref=np.max), sr=sr, hop_length=len_hop, y_axis='log', x_axis='time')
+    librosa.display.specshow(librosa.amplitude_to_db(spec, ref=np.max), sr=sr, hop_length=len_hop, y_axis='linear', x_axis='time')
     plt.title('spectrogram')
     plt.colorbar(format='%+2.0f dB')
 

tools/stat_amplitude.py → tools/statistics_amp.py

@@ -6,7 +6,9 @@
 import librosa
 import numpy as np
 
-src_path = '/Users/avin/git/vc/datasets/IU_split'
+src_path = '/Users/avin/git/vc/datasets/IU/v_app_split'
+# src_path = '/Users/avin/git/vc/datasets/IU/melon_radio_season1_split'
+# src_path = '/Users/avin/git/vc/datasets/IU/melon_radio_season2_split'
 # src_path = '/Users/avin/git/vc/datasets/timit/TIMIT/TRAIN/*/*'
 # src_path = '/Users/avin/git/vc/datasets/kate/sense_and_sensibility_split'
 # src_path = '/Users/avin/git/vc/datasets/arctic/bdl'
@@ -38,26 +40,25 @@
 dbs = np.array(dbs)
 
 
-mean = np.mean(amps)
-std = np.std(amps)
+mean_amps = np.mean(amps)
+std_amps = np.std(amps)
+max_amps = np.max(amps)
+min_amps = np.min(amps)
 
-max = np.max(amps)
-min = np.min(amps)
+mean_dbs = np.mean(dbs)
+std_dbs = np.std(dbs)
+max_dbs = np.max(dbs)
+min_dbs = np.min(dbs)
 
-# mean = np.mean(dbs)
-# std = np.std(dbs)
-#
-# max = np.max(dbs)
-# min = np.min(dbs)
-
-# mean = np.mean(log_amps)
-# std = np.std(log_amps)
-#
-# max = np.max(log_amps)
-# min = np.min(log_amps)
+mean_log_amps = np.mean(log_amps)
+std_log_amps = np.std(log_amps)
+max_log_amps = np.max(log_amps)
+min_log_amps = np.min(log_amps)
 
 # normalized = (values - mean) / std
 # normalized = (log_amps - min) / (max - min)
 
-print("max: {}, min: {}, mean: {}, std: {}".format(max, min, mean, std))
+print("[amps]     max: {}, min: {}, mean: {}, std: {}".format(max_amps, min_amps, mean_amps, std_amps))
+print("[log_amps] max: {}, min: {}, mean: {}, std: {}".format(max_log_amps, min_log_amps, mean_log_amps, std_log_amps))
+print("[decibels] max: {}, min: {}, mean: {}, std: {}".format(max_dbs, min_dbs, mean_dbs, std_dbs))
 # print(normalized)

train2.py

@@ -104,6 +104,8 @@ def get_arguments():
     logdir2 = '{}/{}/train2'.format(logdir_path, case2)
     Hparam(case2).set_as_global_hparam()
 
+    print('case1: {}, case2: {}, logdir1: {}, logdir2: {}'.format(case1, case2, logdir1, logdir2))
+
     train(logdir1=logdir1, logdir2=logdir2)
 
     print("Done")