amplitude is normalized decibel

audio.py

@@ -53,6 +53,14 @@ def inv_preemphasis(x, coeff=0.97):
     return signal.lfilter([1], [1, -coeff], x)
 
 
+def amp_to_db(amp):
+    return librosa.amplitude_to_db(amp)
+
+
+def db_to_amp(db):
+    return librosa.db_to_amplitude(db)
+
+
 def split(wav, top_db):
     intervals = librosa.effects.split(wav, top_db=top_db)
     wavs = map(lambda i: wav[i[0]: i[1]], intervals)

convert.py

@@ -9,11 +9,12 @@
 from data_load import get_wav_batch, get_batch
 from models import Model
 import numpy as np
-from audio import spectrogram2wav, inv_preemphasis
+from audio import spectrogram2wav, inv_preemphasis, db_to_amp
 from hparam import logdir_path
 import datetime
 import tensorflow as tf
 from hparam import Hparam
+from utils import denormalize_0_1
 
 
 def convert(logdir, step, writer, queue=False):
@@ -41,31 +42,37 @@ def convert(logdir, step, writer, queue=False):
         epoch, gs = Model.get_epoch_and_global_step(logdir, step=step)
 
         if queue:
-            pred_log_specs, y_log_spec, ppgs = sess.run([model(), model.y_spec, model.ppgs])
+            pred_spec, y_spec, ppgs = sess.run([model(), model.y_spec, model.ppgs])
         else:
             if hp.convert.one_full_wav:
                 mfcc, spec, mel = get_wav_batch(model.mode, model.batch_size)
             else:
                 mfcc, spec, mel = get_batch(model.mode, model.batch_size)
 
-            pred_log_specs, y_log_spec, ppgs = sess.run([model(), model.y_spec, model.ppgs], feed_dict={model.x_mfcc: mfcc, model.y_spec: spec, model.y_mel: mel})
+            pred_spec, y_spec, ppgs = sess.run([model(), model.y_spec, model.ppgs], feed_dict={model.x_mfcc: mfcc, model.y_spec: spec, model.y_mel: mel})
+
+        # De-quantization
+        # bins = np.linspace(0, 1, hp.default.quantize_db)
+        # y_spec = bins[y_spec]
 
         # Denormalizatoin
-        # pred_log_specs = hp.mean_log_spec + hp.std_log_spec * pred_log_specs
-        # y_log_spec = hp.mean_log_spec + hp.std_log_spec * y_log_spec
-        # pred_log_specs = hp.min_log_spec + (hp.max_log_spec - hp.min_log_spec) * pred_log_specs
-        # y_log_spec = hp.min_log_spec + (hp.max_log_spec - hp.min_log_spec) * y_log_spec
+        pred_spec = denormalize_0_1(pred_spec, hp.default.max_db, hp.default.min_db)
+        y_spec = denormalize_0_1(y_spec, hp.default.max_db, hp.default.min_db)
+
+        # Db to amp
+        pred_spec = db_to_amp(pred_spec)
+        y_spec = db_to_amp(y_spec)
 
         # Convert log of magnitude to magnitude
-        pred_specs, y_specs = np.e ** pred_log_specs, np.e ** y_log_spec
+        # pred_specs, y_specs = np.e ** pred_specs, np.e ** y_spec
 
         # Emphasize the magnitude
-        pred_specs = np.power(pred_specs, hp.convert.emphasis_magnitude)
-        y_specs = np.power(y_specs, hp.convert.emphasis_magnitude)
+        pred_spec = np.power(pred_spec, hp.convert.emphasis_magnitude)
+        y_spec = np.power(y_spec, hp.convert.emphasis_magnitude)
 
         # Spectrogram to waveform
-        audio = np.array(map(lambda spec: spectrogram2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.default.hop_length, hp.default.n_iter), pred_specs))
-        y_audio = np.array(map(lambda spec: spectrogram2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.default.hop_length, hp.default.n_iter), y_specs))
+        audio = np.array(map(lambda spec: spectrogram2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.default.hop_length, hp.default.n_iter), pred_spec))
+        y_audio = np.array(map(lambda spec: spectrogram2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.default.hop_length, hp.default.n_iter), y_spec))
 
         # Apply inverse pre-emphasis
         audio = inv_preemphasis(audio, coeff=hp.default.preemphasis)

data_load.py

@@ -2,7 +2,6 @@
 # /usr/bin/python2
 
 import glob
-import sys
 import threading
 from functools import wraps
 from random import sample
@@ -11,10 +10,11 @@
 from tensorflow.python.platform import tf_logging as logging
 
 from hparam import Hparam
-from audio import preemphasis
+from audio import preemphasis, amp_to_db
 import numpy as np
 import librosa
 from hparam import data_path_base
+from utils import normalize_0_1
 
 
 def wav_random_crop(wav, sr, duration):
@@ -119,19 +119,21 @@ def _get_mfcc_log_spec_and_log_mel_spec(wav, preemphasis_coeff, n_fft, win_lengt
     mel_basis = librosa.filters.mel(hp.default.sr, hp.default.n_fft, hp.default.n_mels)  # (n_mels, 1+n_fft//2)
     mel = np.dot(mel_basis, mag)  # (n_mels, t) # mel spectrogram
 
-    # Get mfccs
-    db = librosa.amplitude_to_db(mel)
-    mfccs = np.dot(librosa.filters.dct(hp.default.n_mfcc, db.shape[0]), db)
+    # Get mfccs, amp to db
+    mag_db = amp_to_db(mag)
+    mel_db = amp_to_db(mel)
+    mfccs = np.dot(librosa.filters.dct(hp.default.n_mfcc, mel_db.shape[0]), mel_db)
 
-    # Log
-    mag = np.log(mag + sys.float_info.epsilon)
-    mel = np.log(mel + sys.float_info.epsilon)
+    # Normalization (0 ~ 1)
+    mag_db = normalize_0_1(mag_db, hp.default.max_db, hp.default.min_db)
+    mel_db = normalize_0_1(mel_db, hp.default.max_db, hp.default.min_db)
 
-    # Normalization
-    # self.y_log_spec = (y_log_spec - hp.mean_log_spec) / hp.std_log_spec
-    # self.y_log_spec = (y_log_spec - hp.min_log_spec) / (hp.max_log_spec - hp.min_log_spec)
+    # Quantization
+    # bins = np.linspace(0, 1, hp.default.quantize_db)
+    # mag_db = np.digitize(mag_db, bins)
+    # mel_db = np.digitize(mel_db, bins)
 
-    return mfccs.T, mag.T, mel.T  # (t, n_mfccs), (t, 1+n_fft/2), (t, n_mels)
+    return mfccs.T, mag_db.T, mel_db.T  # (t, n_mfccs), (t, 1+n_fft/2), (t, n_mels)
 
 
 # Adapted from the `sugartensor` code.

hparams/default.yaml

@@ -11,6 +11,9 @@ default:
     n_iter: 60 # Number of inversion iterations
     n_mels: 80
     duration: 2
+    max_db: 35
+    min_db: -55
+    quantize_db: 101
 
     # model
     hidden_units: 256  # alias: E

utils.py

@@ -3,6 +3,7 @@
 
 import os
 import glob
+import numpy as np
 
 
 def split_path(path):
@@ -20,3 +21,13 @@ def remove_all_files(prefix):
     files = glob.glob(prefix + '*')
     for f in files:
         os.remove(f)
+
+
+def normalize_0_1(values, max, min):
+    normalized = np.clip((values - min) / (max - min), 0, 1)
+    return normalized
+
+
+def denormalize_0_1(normalized, max, min):
+    values =  np.clip(normalized, 0, 1) * (max - min) + min
+    return values