diff --git a/tools/audio_utils.py b/audio.py
similarity index 69%
rename from tools/audio_utils.py
rename to audio.py
index 75278314..9ab80817 100644
--- a/tools/audio_utils.py
+++ b/audio.py
@@ -1,9 +1,5 @@
 # -*- coding: utf-8 -*-
 #!/usr/bin/env python
-'''
-By Dabi Ahn. andabi412@gmail.com.
-https://www.github.com/andabi
-'''
 
 from pydub import AudioSegment
 import os
@@ -34,29 +30,7 @@ def rewrite_mp3_to_wav(source_path, target_path):
     AudioSegment.from_mp3(source_path).export(target_path, format='wav')
 
 
-def split_path(path):
-    '''
-    'a/b/c.wav' => ('a/b', 'c', 'wav')
-    :param path: filepath = 'a/b/c.wav'
-    :return: basename, filename, and extension = ('a/b', 'c', 'wav')
-    '''
-    basepath, filename = os.path.split(path)
-    filename, extension = os.path.splitext(filename)
-    return basepath, filename, extension
-
-
 def spectrogram2wav(mag, n_fft, win_length, hop_length, num_iters, phase_angle=None, length=None):
-    '''
-
-    :param mag: [f, t]
-    :param n_fft: n_fft
-    :param win_length: window length
-    :param hop_length: hop length
-    :param num_iters: num of iteration when griffin-lim reconstruction
-    :param phase_angle: phase angle
-    :param length: length of wav
-    :return: 
-    '''
     assert (num_iters > 0)
     if phase_angle is None:
         phase_angle = np.pi * np.random.rand(*mag.shape)
@@ -77,3 +51,9 @@ def preemphasis(x, coeff=0.97):
 
 def inv_preemphasis(x, coeff=0.97):
     return signal.lfilter([1], [1, -coeff], x)
+
+
+def split(wav, top_db):
+    intervals = librosa.effects.split(wav, top_db=top_db)
+    wavs = map(lambda i: wav[i[0]: i[1]], intervals)
+    return wavs
\ No newline at end of file
diff --git a/convert.py b/convert.py
index b56ae103..0a9209c4 100644
--- a/convert.py
+++ b/convert.py
@@ -9,7 +9,7 @@
 from data_load import get_wav_batch, get_batch
 from models import Model
 import numpy as np
-from utils import spectrogram2wav, inv_preemphasis
+from audio import spectrogram2wav, inv_preemphasis
 from hparam import logdir_path
 import datetime
 import tensorflow as tf
diff --git a/data_load.py b/data_load.py
index 37b6ad19..dbe937f5 100644
--- a/data_load.py
+++ b/data_load.py
@@ -11,12 +11,26 @@
 from tensorflow.python.platform import tf_logging as logging
 
 from hparam import Hparam
-from utils import preemphasis, wav_random_crop
+from audio import preemphasis
 import numpy as np
 import librosa
 from hparam import data_path_base
 
 
+def wav_random_crop(wav, sr, duration):
+    assert (wav.ndim <= 2)
+
+    target_len = sr * duration
+    wav_len = wav.shape[-1]
+    start = np.random.choice(range(np.maximum(1, wav_len - target_len)), 1)[0]
+    end = start + target_len
+    if wav.ndim == 1:
+        wav = wav[start:end]
+    else:
+        wav = wav[:, start:end]
+    return wav
+
+
 def get_mfccs_and_phones(wav_file, sr, length, trim=False, random_crop=True):
     hp = Hparam.get_global_hparam()
 
diff --git a/hparam.py b/hparam.py
index 570d0dbe..310e0637 100644
--- a/hparam.py
+++ b/hparam.py
@@ -7,11 +7,11 @@
 # path
 ## local
 data_path_base = './datasets'
-logdir_path = './logdir'
+# logdir_path = './logdir'
 
 ## remote
 # data_path_base = '/data/private/vc/datasets'
-# logdir_path = '/data/private/vc/logdir'
+logdir_path = '/data/private/vc/logdir'
 
 
 
@@ -73,6 +73,7 @@ def __call__(self):
 
     def set_as_global_hparam(self):
         Hparam.global_hparam = self.hparam
+        return Hparam.global_hparam
 
     @staticmethod
     def get_global_hparam():
diff --git a/hparams/hparams.yaml b/hparams/hparams.yaml
deleted file mode 100644
index 2b4d60b7..00000000
--- a/hparams/hparams.yaml
+++ /dev/null
@@ -1,15 +0,0 @@
----
-default_iu3:
-  train2:
-      data_path: 'IU/*_split/*.wav'
-  convert:
-      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
\ No newline at end of file
diff --git a/models.py b/models.py
index 35e54e3f..58635f64 100644
--- a/models.py
+++ b/models.py
@@ -108,6 +108,7 @@ def _net2(self):
             # CBHG1: mel-scale
             pred_mel = cbhg(prenet_out, self.hp.train2.num_banks, self.hp.train2.hidden_units // 2, self.hp.train2.num_highway_blocks, self.hp.train2.norm_type, self.is_training, scope="cbhg1")
             pred_mel = tf.layers.dense(pred_mel, self.y_mel.shape[-1])  # log magnitude: (N, T, n_mels)
+            # pred_mel = prenet_out
 
             # CBHG2: linear-scale
             pred_spec = tf.layers.dense(pred_mel, self.hp.train2.hidden_units // 2)  # log magnitude: (N, T, n_mels)
@@ -119,6 +120,7 @@ def _net2(self):
     def loss_net2(self):
         loss_spec = tf.reduce_mean(tf.squared_difference(self.pred_spec, self.y_spec))
         loss_mel = tf.reduce_mean(tf.squared_difference(self.pred_mel, self.y_mel))
+        # loss_mel = 0
         loss = loss_spec + loss_mel
         return loss
 
diff --git a/scripts/tensorboard_start.sh b/scripts/tensorboard.sh
similarity index 100%
rename from scripts/tensorboard_start.sh
rename to scripts/tensorboard.sh
diff --git a/tools/comparison_wavs.py b/tools/comparison_wavs.py
deleted file mode 100644
index effc81db..00000000
--- a/tools/comparison_wavs.py
+++ /dev/null
@@ -1,55 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa.display
-import utils
-import numpy as np
-import matplotlib.pyplot as plt
-
-filename_a = '/Users/avin/git/vc/a.wav'
-filename_b = '/Users/avin/git/vc/b.wav'
-
-n_fft = 4096
-len_hop = n_fft / 4
-plot_wav = True
-plot_spec = True
-
-# Waveforms
-wav_a, sr = librosa.load(filename_a, mono=True)
-wav_b, _ = librosa.load(filename_b, mono=True)
-
-# Spectrogram
-spec_a = librosa.stft(wav_a, n_fft=n_fft, hop_length=len_hop)
-spec_b = librosa.stft(wav_b, n_fft=n_fft, hop_length=len_hop)
-
-# Plot waveforms
-if plot_wav:
-    plt.figure(1)
-
-    plt.subplot(2, 1, 1)
-    librosa.display.waveplot(wav_a, sr=sr, color='b')
-    plt.title('A')
-
-    plt.subplot(2, 1, 2)
-    librosa.display.waveplot(wav_b, sr=sr, color='r')
-    plt.title('B')
-
-    plt.tight_layout()
-    plt.show()
-
-# Plot spectrogram
-if plot_spec:
-    plt.figure(2)
-
-    plt.subplot(2, 1, 1)
-    librosa.display.specshow(librosa.amplitude_to_db(spec_a, ref=np.max), sr=sr, hop_length=len_hop, y_axis='log', x_axis='time')
-    plt.title('A')
-    plt.colorbar(format='%+2.0f dB')
-
-    plt.subplot(2, 1, 2)
-    librosa.display.specshow(librosa.amplitude_to_db(spec_b, ref=np.max), sr=sr, hop_length=len_hop, y_axis='log', x_axis='time')
-    plt.title('B')
-    plt.colorbar(format='%+2.0f dB')
-
-    plt.tight_layout()
-    plt.show()
\ No newline at end of file
diff --git a/tools/convert_amp.py b/tools/convert_amp.py
deleted file mode 100644
index c15d88ec..00000000
--- a/tools/convert_amp.py
+++ /dev/null
@@ -1,25 +0,0 @@
-# -*- 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')
\ No newline at end of file
diff --git a/tools/griffin_lim_recon_test.py b/tools/griffin_lim_recon_test.py
deleted file mode 100644
index 46e55217..00000000
--- a/tools/griffin_lim_recon_test.py
+++ /dev/null
@@ -1,31 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-'''
-By Dabi Ahn. andabi412@gmail.com.
-https://www.github.com/andabi
-'''
-
-import librosa
-import numpy as np
-from audio_utils import spectrogram2wav as griffin_lim
-
-file_path = '/Users/avin/git/vc/datasets/kate/sense_and_sensibility/1-1-001.wav'
-sr = 22050
-n_fft = 512
-win_length = 400
-hop_length = win_length / 4
-num_iters = 30
-
-wav, sr = librosa.load(file_path, sr=sr, mono=True)
-spec = librosa.stft(wav, n_fft=n_fft, hop_length=hop_length)
-mag, phase = np.abs(spec), np.angle(spec)
-
-# Griffin Lim reconstruction
-wav_recon = griffin_lim(mag, n_fft=n_fft, win_length=win_length, hop_length=hop_length, num_iters=num_iters, length=wav.shape[0])
-# sum_wav = np.sum(np.abs(wav))
-# sum_wav_recon = np.sum(np.abs(wav_recon))
-# print(sum_wav, sum_wav_recon, (sum_wav - sum_wav_recon) / sum_wav * 100)
-
-# Write
-librosa.output.write_wav('wav_orig.wav', wav, sr)
-librosa.output.write_wav('wav_recon_{}.wav'.format(num_iters), wav_recon, sr)
\ No newline at end of file
diff --git a/tools/librosa_load_test.py b/tools/librosa_load_test.py
deleted file mode 100644
index aef99a83..00000000
--- a/tools/librosa_load_test.py
+++ /dev/null
@@ -1,21 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa
-import datetime
-
-filename = '/Users/avin/git/vc/datasets/timit/TIMIT/TRAIN/DR1/FCJF0/SA1.wav'  # sample rate = 16,000 Hz
-
-# same sample rate
-s = datetime.datetime.now()
-wav, _ = librosa.load(filename, mono=True, sr=16000)
-e = datetime.datetime.now()
-diff = e - s
-print(diff.microseconds)
-
-# different sample rate (22,050 Hz)
-s = datetime.datetime.now()
-wav, _ = librosa.load(filename, mono=True)
-e = datetime.datetime.now()
-diff = e - s
-print(diff.microseconds)
diff --git a/tools/librosa_split_test.py b/tools/librosa_split_test.py
deleted file mode 100644
index 902c6d0d..00000000
--- a/tools/librosa_split_test.py
+++ /dev/null
@@ -1,14 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa
-import datetime
-
-filename = '/Users/avin/git/vc/datasets/kate/sense_and_sensibility/1-1-003.wav'
-
-wav, sr = librosa.load(filename, mono=True, sr=22050)
-intervals = librosa.effects.split(wav, top_db=30)
-for i, interval in enumerate(intervals):
-    s, e = interval[0], interval[1]
-    w = wav[s:e]
-    librosa.output.write_wav('/Users/avin/git/vc/{}.wav'.format(i), w, sr)
diff --git a/tools/mfcc_amp_normalization_test.py b/tools/mfcc_amp_normalization_test.py
deleted file mode 100644
index a6099fdf..00000000
--- a/tools/mfcc_amp_normalization_test.py
+++ /dev/null
@@ -1,49 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa
-import numpy as np
-
-filename = '/Users/avin/git/vc/datasets/timit/TIMIT/TRAIN/DR1/FCJF0/SA1.wav'  # sample rate = 16,000 Hz
-
-sr = 16000
-n_fft = 512
-win_length = 400
-hop_length = 80
-n_mels = 128
-n_mfcc = 40
-
-# Load waveforms
-y, _ = librosa.load(filename, mono=True, sr=sr)
-
-# Get spectrogram
-D = librosa.stft(y=y,
-                 n_fft=n_fft,
-                 hop_length=hop_length,
-                 win_length=win_length)
-mag = np.abs(D)  # (1+n_fft//2, t)
-scaled_mag = mag * 200
-
-# Get mel-spectrogram
-mel_basis = librosa.filters.mel(sr, n_fft, n_mels)  # (n_mels, 1+n_fft//2)
-mel = np.dot(mel_basis, mag)  # (n_mels, t) # mel spectrogram
-scaled_mel = np.dot(mel_basis, scaled_mag)
-
-# Get mfccs
-db = librosa.amplitude_to_db(mel)
-scaled_db = librosa.amplitude_to_db(scaled_mel)
-
-print(db)
-print(scaled_db)
-
-mfccs = np.dot(librosa.filters.dct(n_mfcc, db.shape[0]), db)
-scaled_mfccs = np.dot(librosa.filters.dct(n_mfcc, db.shape[0]), scaled_db)
-
-mfccs = mfccs.T  # (t, n_mfccs)
-scaled_mfccs = scaled_mfccs.T
-
-# assert(np.all(mfccs * 2 == scaled_mfccs))
-
-print(mfccs / scaled_mfccs)
-print(mfccs)
-print(scaled_mfccs)
\ No newline at end of file
diff --git a/tools/mp3_to_wav.py b/tools/mp3_to_wav.py
deleted file mode 100644
index 21f94261..00000000
--- a/tools/mp3_to_wav.py
+++ /dev/null
@@ -1,14 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import glob
-import audio_utils
-
-SOURCE_PATH = 'datasets/test/godfather2'
-TARGET_PATH = 'datasets/test/godfather2'
-
-# Write mp3 to wav
-search_path = '{}/*.mp3'.format(SOURCE_PATH)
-for source_path in glob.glob(search_path):
-    target_path = '{}/{}.wav'.format(SOURCE_PATH, 'a')
-    audio_utils.rewrite_mp3_to_wav(TARGET_PATH, target_path)
diff --git a/tools/plot_spectrogram.py b/tools/plot_spectrogram.py
deleted file mode 100644
index ca3d5b46..00000000
--- a/tools/plot_spectrogram.py
+++ /dev/null
@@ -1,45 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa.display
-import utils
-import numpy as np
-import matplotlib.pyplot as plt
-from audio_utils import read, write
-
-filename = '/Users/avin/git/vc/outputs/male.wav'
-sr = 22050
-n_fft = 1024
-len_hop = n_fft / 4
-plot_wav = False
-plot_spec = True
-
-# Waveforms
-wav = read(filename, sr, mono=True)
-# wav = np.where(wav == 0, 1000, wav)
-# wav = np.zeros_like(wav)
-# wav[0] = np.ones_like(wav[0])
-
-# Spectrogram
-spec = librosa.stft(wav, n_fft=n_fft, hop_length=len_hop)
-
-# Plot waveforms
-if plot_wav:
-    plt.figure(1)
-
-    librosa.display.waveplot(wav, sr=sr, color='b')
-    plt.title('waveform')
-
-    plt.tight_layout()
-    plt.show()
-
-# Plot spectrogram
-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='linear', x_axis='time')
-    plt.title('spectrogram')
-    plt.colorbar(format='%+2.0f dB')
-
-    plt.tight_layout()
-    plt.show()
\ No newline at end of file
diff --git a/tools/preemphasis_test.py b/tools/preemphasis_test.py
deleted file mode 100644
index f52066fd..00000000
--- a/tools/preemphasis_test.py
+++ /dev/null
@@ -1,63 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa.display
-import numpy as np
-import matplotlib.pyplot as plt
-from audio_utils import read, write, preemphasis, inv_preemphasis
-import sys
-
-filename = '/Users/avin/git/vc/datasets/timit/TIMIT/TEST/DR1/FAKS0/SA1.wav'
-sr = 22050
-n_fft = 4096
-len_hop = n_fft / 4
-write_wav = True
-plot_wav = True
-plot_spec = True
-
-# Waveforms
-wav = read(filename, sr, mono=True)
-wav_empha = preemphasis(wav)
-diff = wav - inv_preemphasis(wav_empha)
-assert(np.all(np.less(diff, sys.float_info.epsilon)))
-
-# Spectrogram
-spec = librosa.stft(wav, n_fft=n_fft, hop_length=len_hop)
-spec_empha = librosa.stft(wav_empha, n_fft=n_fft, hop_length=len_hop)
-
-# Write wav
-if write_wav:
-    write(wav, sr, 'original2.wav')
-    write(wav_empha, sr, 'preemphasized2.wav')
-
-# Plot waveforms
-if plot_wav:
-    plt.figure(1)
-
-    plt.subplot(2, 1, 1)
-    librosa.display.waveplot(wav, sr=sr, color='b')
-    plt.title('before preemphasis')
-
-    plt.subplot(2, 1, 2)
-    librosa.display.waveplot(wav_empha, sr=sr, color='r')
-    plt.title('after preemphasis')
-
-    plt.tight_layout()
-    plt.show()
-
-# Plot spectrogram
-if plot_spec:
-    plt.figure(2)
-
-    plt.subplot(2, 1, 1)
-    librosa.display.specshow(librosa.amplitude_to_db(spec, ref=np.max), sr=sr, hop_length=len_hop, y_axis='log', x_axis='time')
-    plt.title('before preemphasis')
-    plt.colorbar(format='%+2.0f dB')
-
-    plt.subplot(2, 1, 2)
-    librosa.display.specshow(librosa.amplitude_to_db(spec_empha, ref=np.max), sr=sr, hop_length=len_hop, y_axis='log', x_axis='time')
-    plt.title('after preemphasis')
-    plt.colorbar(format='%+2.0f dB')
-
-    plt.tight_layout()
-    plt.show()
\ No newline at end of file
diff --git a/tools/preprocess_wav.py b/tools/preprocess_wav.py
deleted file mode 100644
index a9e7e88c..00000000
--- a/tools/preprocess_wav.py
+++ /dev/null
@@ -1,57 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa
-import glob
-import os
-from audio_utils import read, write, split_path
-
-
-# src_path = '/Users/avin/git/vc/datasets/arctic/bdl'
-src_path = '/Users/avin/git/vc/datasets/IU/melon_radio_season1'
-top_db = 30
-target_path = '{}_split_{}'.format(src_path, top_db)
-sr = 16000
-min_limit_len = sr / 2  # 0.5s
-
-# Preprocessing
-# 1. split wav by silence
-# 2. convert sample rate
-
-
-def split(wav, top_db):
-    intervals = librosa.effects.split(wav, top_db=top_db)
-    wavs = map(lambda i: wav[i[0]: i[1]], intervals)
-    return wavs
-
-
-num_files = num_files_split = 0
-max_len = max_len_split = 0
-min_len = min_len_split = float('inf')
-for filepath in glob.glob('{}/*.wav'.format(src_path)):
-    wav = read(filepath, sr, mono=True)
-    split_wavs = split(wav, top_db)
-    if not os.path.exists(target_path):
-        os.mkdir(target_path)
-    _, filename, _ = split_path(filepath)
-
-    # Statistics
-    # num_files += 1
-    # num_files_split += len(split_wavs)
-    #
-    # max_len = max(max_len, len(wav))
-    # max_len_split = max(max_len_split, max(map(lambda w: len(w), split_wavs)))
-    #
-    # min_len = min(min_len, len(wav))
-    # min_len_split = min(min_len_split, min(map(lambda w: len(w), split_wavs)))
-
-    split_wavs = filter(lambda w: len(w) >= min_limit_len, split_wavs)
-    map(lambda (i, w): write(w, sr, '{}/{}_{}.wav'.format(target_path, filename, i)), enumerate(split_wavs))
-
-
-# print('num_files: {}, num_files_split: {}'.format(num_files, num_files_split))
-# print('max_len: {}, max_len_split: {}'.format(max_len, max_len_split))
-# print('min_len: {}, min_len_split: {}'.format(min_len, min_len_split))
-print('done')
-
-# range of arctic/bdl: 20000~120000
\ No newline at end of file
diff --git a/tools/split_wav.py b/tools/split_wav.py
deleted file mode 100644
index f6181158..00000000
--- a/tools/split_wav.py
+++ /dev/null
@@ -1,53 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import librosa
-import glob
-import os
-from audio_utils import read, write, split_path
-
-
-# src_path = '/Users/avin/git/vc/datasets/arctic/bdl'
-src_path = '/Users/avin/git/vc/datasets/IU'
-target_path = '{}_split'.format(src_path)
-sr = 16000
-top_db = 25
-min_limit_len = sr  # 1s
-
-
-def split(wav, top_db):
-    intervals = librosa.effects.split(wav, top_db=top_db)
-    wavs = map(lambda i: wav[i[0]: i[1]], intervals)
-    return wavs
-
-
-num_files = num_files_split = 0
-max_len = max_len_split = 0
-min_len = min_len_split = float('inf')
-for filepath in glob.glob('{}/**/*.wav'.format(src_path)):
-    wav = read(filepath, sr, mono=True)
-    split_wavs = split(wav, top_db)
-    if not os.path.exists(target_path):
-        os.mkdir(target_path)
-    _, filename, _ = split_path(filepath)
-
-    # Statistics
-    # num_files += 1
-    # num_files_split += len(split_wavs)
-    #
-    # max_len = max(max_len, len(wav))
-    # max_len_split = max(max_len_split, max(map(lambda w: len(w), split_wavs)))
-    #
-    # min_len = min(min_len, len(wav))
-    # min_len_split = min(min_len_split, min(map(lambda w: len(w), split_wavs)))
-
-    split_wavs = filter(lambda w: len(w) >= min_limit_len, split_wavs)
-    map(lambda (i, w): write(w, sr, '{}/{}_{}.wav'.format(target_path, filename, i)), enumerate(split_wavs))
-
-
-# print('num_files: {}, num_files_split: {}'.format(num_files, num_files_split))
-# print('max_len: {}, max_len_split: {}'.format(max_len, max_len_split))
-# print('min_len: {}, min_len_split: {}'.format(min_len, min_len_split))
-print('done')
-
-# range of arctic/bdl: 20000~120000
\ No newline at end of file
diff --git a/tools/statistics_amp.py b/tools/statistics_amp.py
deleted file mode 100644
index 7ce70a20..00000000
--- a/tools/statistics_amp.py
+++ /dev/null
@@ -1,64 +0,0 @@
-# -*- coding: utf-8 -*-
-#!/usr/bin/env python
-
-import glob
-from audio_utils import read
-import librosa
-import numpy as np
-
-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'
-# src_path = '/Users/avin/git/vc/datasets/arctic/slt'
-# src_path = '/Users/avin/git/vc/datasets/kate/therese_raquin_split'
-sr = 16000
-n_fft = 512
-win_length = 400
-hop_length = 80
-n_sample = 200
-
-amps = []
-log_amps = []
-dbs = []
-for filepath in glob.glob('{}/*.wav'.format(src_path))[:n_sample]:
-    wav = read(filepath, sr, mono=True)
-    spec = librosa.stft(wav, n_fft=n_fft, win_length=win_length, hop_length=hop_length)  # (n_fft/2+1, t)
-    amp = np.abs(spec)
-    amps.extend(amp.flatten())
-
-    log_amp = np.log(amp)
-    log_amps.extend(log_amp.flatten())
-
-    db = librosa.amplitude_to_db(amp)
-    dbs.extend(db.flatten())
-
-amps = np.array(amps)
-log_amps = np.array(log_amps)
-dbs = np.array(dbs)
-
-
-mean_amps = np.mean(amps)
-std_amps = np.std(amps)
-max_amps = np.max(amps)
-min_amps = np.min(amps)
-
-mean_dbs = np.mean(dbs)
-std_dbs = np.std(dbs)
-max_dbs = np.max(dbs)
-min_dbs = np.min(dbs)
-
-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("[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)
\ No newline at end of file