def convert(predictor, df):
a, b, c = next(df().get_data())
pred_spec, r_spec = predictor(a, b, c)
# Denormalizatoin
pred_spec = denormalize_db(pred_spec, hp.default.max_db, hp.default.min_db)
r_spec = denormalize_db(r_spec, hp.default.max_db, hp.default.min_db)
# Db to amp
pred_spec = db2amp(pred_spec)
r_spec = db2amp(r_spec)
# Emphasize the magnitude
pred_spec = np.power(pred_spec, hp.convert.emphasis_magnitude)
r_spec = np.power(r_spec, hp.convert.emphasis_magnitude)
# Spectrogram to waveform
audio = np.array(
list(map(lambda spec: spec2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.default.hop_length,
hp.default.n_iter), pred_spec)))
y_audio = np.array(
list(map(lambda spec: spec2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.default.hop_length,
hp.default.n_iter), r_spec)))
# Apply inverse pre-emphasis
audio = inv_preemphasis(audio, coeff=hp.default.preemphasis)
y_audio = inv_preemphasis(y_audio, coeff=hp.default.preemphasis)
# if hp.convert.one_full_wav:
# # Concatenate to a wav
# y_audio = np.reshape(y_audio, (1, y_audio.size), order='C')
# audio = np.reshape(audio, (1, audio.size), order='C')
return audio, y_audio
def convert(predictor, df):
pred_spec, y_spec, ppgs = predictor(next(df().get_data()))
# Denormalizatoin
pred_spec = denormalize_db(pred_spec, hp.Default.max_db, hp.Default.min_db)
y_spec = denormalize_db(y_spec, hp.Default.max_db, hp.Default.min_db)
# Db to amp
pred_spec = db2amp(pred_spec)
y_spec = db2amp(y_spec)
# Emphasize the 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([spec2wav(spec.T, hp.Default.n_fft, hp.Default.win_length, hp.Default.hop_length,
hp.Default.n_iter) for spec in pred_spec])
y_audio = np.array([spec2wav(spec.T, hp.Default.n_fft, hp.Default.win_length, hp.Default.hop_length,
hp.Default.n_iter) for spec in y_spec])
# Apply inverse pre-emphasis
audio = inv_preemphasis(audio, coeff=hp.Default.preemphasis)
y_audio = inv_preemphasis(y_audio, coeff=hp.Default.preemphasis)
# if hp.Convert.one_full_wav:
# # Concatenate to a wav
# y_audio = np.reshape(y_audio, (1, y_audio.size), order='C')
# audio = np.reshape(audio, (1, audio.size), order='C')
return audio, y_audio, ppgs
def convert(predictor, df):
t = next(df().get_data())
print(t[0].shape)
pred_spec, y_spec, ppgs = predictor(next(df().get_data()))
# Denormalizatoin
pred_spec = denormalize_db(pred_spec, hp.default.max_db, hp.default.min_db)
y_spec = denormalize_db(y_spec, hp.default.max_db, hp.default.min_db)
# Db to amp
pred_spec = db2amp(pred_spec)
y_spec = db2amp(y_spec)
# Emphasize the 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:
spec2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.
default.hop_length, hp.default.n_iter), pred_spec))
librosa.output.write_wav(
'/home/user/vilin/deep-voice-conversion/output/file_trim_8.wav',
audio[0], hp.default.sr)
y_audio = np.array(
map(
lambda spec:
spec2wav(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)
y_audio = inv_preemphasis(y_audio, coeff=hp.default.preemphasis)
# if hp.convert.one_full_wav:
# # Concatenate to a wav
# y_audio = np.reshape(y_audio, (1, y_audio.size), order='C')
# audio = np.reshape(audio, (1, audio.size), order='C')
return audio, y_audio, ppgs
def convert(predictor, data):
x_mfccs, y_spec, y_mel = data
x_mfccs, y_spec, y_mel = data
x_mfccs = np.array(x_mfccs).reshape((-1, ) + x_mfccs.shape)
y_spec = np.array(y_spec).reshape((-1, ) + y_spec.shape)
y_mel = np.array(y_mel).reshape((-1, ) + y_mel.shape)
pred_spec, y_spec, ppgs = predictor(x_mfccs, y_spec, y_mel)
# Denormalizatoin
pred_spec = denormalize_db(pred_spec, hp.Default.max_db, hp.Default.min_db)
y_spec = denormalize_db(y_spec, hp.Default.max_db, hp.Default.min_db)
# Db to amp
pred_spec = db2amp(pred_spec)
y_spec = db2amp(y_spec)
# Emphasize the 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([
spec2wav(spec.T, hp.Default.n_fft, hp.Default.win_length,
hp.Default.hop_length, hp.Default.n_iter)
for spec in pred_spec
])
y_audio = np.array([
spec2wav(spec.T, hp.Default.n_fft, hp.Default.win_length,
hp.Default.hop_length, hp.Default.n_iter) for spec in y_spec
])
# Apply inverse pre-emphasis
audio = inv_preemphasis(audio, coeff=hp.Default.preemphasis)
y_audio = inv_preemphasis(y_audio, coeff=hp.Default.preemphasis)
if hp.Convert.one_full_wav:
# Concatenate to a wav
y_audio = np.reshape(y_audio, (1, y_audio.size), order='C')
audio = np.reshape(audio, (1, audio.size), order='C')
return audio, y_audio, ppgs
def predict(self, path_to_wav):
x_mfcc, y_spec, mel = get_mfccs_and_spectrogram(path_to_wav)
pred_spec, y_spec, ppgs = self.predictor(x_mfcc, y_spec, mel)
pred_spec = denormalize_db(pred_spec, hp.default.max_db,
hp.default.min_db)
y_spec = denormalize_db(y_spec, hp.default.max_db, hp.default.min_db)
pred_spec = db2amp(pred_spec)
y_spec = db2amp(y_spec)
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:
spec2wav(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:
spec2wav(spec.T, hp.default.n_fft, hp.default.win_length, hp.
default.hop_length, hp.default.n_iter), y_spec))
def convert(predictor, df):
# TODO need to fix reading in with duration
pred_spec, y_spec, ppgs = predictor(next(df().get_data()))
# Denormalizatoin
pred_spec = denormalize_db(pred_spec, hp.default.max_db, hp.default.min_db)
y_spec = denormalize_db(y_spec, hp.default.max_db, hp.default.min_db)
# Db to amp
pred_spec = db2amp(pred_spec)
y_spec = db2amp(y_spec)
# Emphasize the 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:
spec2wav(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:
spec2wav(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)
y_audio = inv_preemphasis(y_audio, coeff=hp.default.preemphasis)
if hp.convert.one_full_wav:
# Concatenate to a wav
y_audio = np.reshape(y_audio, (1, y_audio.size), order='C')
audio = np.reshape(audio, (1, audio.size), order='C')
return audio, y_audio, ppgs
def sumimage(mel, mel_name):
mel = mel #+ 0.001 * np.random.standard_normal([hp.batch_size, hp.duration * hp.n_mels, hp.n_mels])
mel_image = mel.transpose(0, 2, 1)
heatmap = np.expand_dims(mel_image, 3)
tf.summary.image(mel_name, heatmap, max_outputs=mel_image.shape[0])
mel_basis = librosa.filters.mel(hp.sr, hp.n_fft, hp.n_mels)
mel_basis = np.mat(mel_basis)
mel_basis_I = mel_basis.I
mel_spec = []
for i in range(len(mel)):
print(mel_name)
print(np.max(mel[i]))
print(np.min(mel[i]))
print(np.mean(mel[i]))
#mel[i] = mel[i] * (0.6 / np.max(mel[i]))
mel_db_item = np.transpose(mel[i])
mel_db_item = denormalize_0_1(mel_db_item, hp.max_db, hp.min_db)
#mel_db_item = np.maximum(mel_db_item, 0)
# = normalize_0_1(mel_db_item, hp.default.max_db, hp.default.min_db)
print(np.max(mel_db_item))
print(np.mean(mel_db_item))
mel_item = db2amp(mel_db_item)
print(np.max(mel_item))
mag_item = np.dot(mel_basis_I, mel_item)
print(np.max(mel_item))
mag_item = np.maximum(mag_item, 0)
spec_item = np.transpose(mag_item)
#mag_db_item = amp2db(mag_item)
#mag_db_item = normalize_0_1(mag_db_item, hp.default.max_db, hp.default.min_db)
#mag_db_item = np.transpose(mag_db_item)
#specitem = np.transpose(magitem)
#mel_complex = mel_D_abs + np.complex(0, 0)
#specitem = librosa.istft(stft_matrix=mel_complex, hop_length=hp.default.hop_length, win_length=hp.default.win_length)
mel_spec.append(spec_item.getA())
mel_spec = np.power(mel_spec, hp.emphasis_magnitude)
mel_audio = np.array(
list(
map(
lambda spec: spec2wav(spec.T, hp.n_fft, hp.win_length, hp.
hop_length, hp.n_iter), mel_spec)))
mel_audio = inv_preemphasis(mel_audio, coeff=hp.preemphasis)
tf.summary.audio(mel_name, mel_audio, hp.sr, max_outputs=hp.batch_size)
def sumspecimage(spec, spec_name):
spec = denormalize_db(spec, hp.max_db, hp.min_db)
spec = db2amp(spec)
spec_image = spec.transpose(0, 2, 1)
heatmap = np.expand_dims(spec_image, 3)
tf.summary.image(spec_name, heatmap, max_outputs=spec_image.shape[0])
out_spec = np.power(np.maximum(spec, 0), 1) #hp.emphasis_magnitude)
out_audio = np.array(
list(
map(
lambda spec: spec2wav(spec.T, hp.n_fft, hp.win_length, hp.
hop_length, hp.n_iter), out_spec)))
out_audio = inv_preemphasis(out_audio, coeff=hp.preemphasis)
tf.summary.audio(spec_name, out_audio, hp.sr, max_outputs=hp.batch_size)
def convert(predictor, tensor):
# tensor = next(df().get_data())
# print(tensor.shape)
pred_spec, y_spec, ppgs = predictor(tensor)
# pred_spec, y_spec, ppgs = predictor(tf.expand_dims(df, 0))
# Denormalizatoin
pred_spec = denormalize_db(pred_spec, hp.default.max_db, hp.default.min_db)
# y_spec = denormalize_db(y_spec, hp.default.max_db, hp.default.min_db)
# Db to amp
pred_spec = db2amp(pred_spec)
# y_spec = db2amp(y_spec)
# Emphasize the 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:
spec2wav(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: spec2wav(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)
# y_audio = inv_preemphasis(y_audio, coeff=hp.default.preemphasis)
# pickle.dump( y_audio, open( "y-audio.p", "wb" ) )
# pickle.dump( audio, open( "o-audio.p", "wb" ) )
# if hp.convert.one_full_wav:
# # Concatenate to a wav
# y_audio = np.reshape(y_audio, (1, y_audio.size), order='C')
# audio = np.reshape(audio, (1, audio.size), order='C')
# return audio, y_audio, ppgs
return audio, ppgs
import numpy as np
import matplotlib.pyplot as plt
from audio import spec2wav, wav2spec, read_wav, write_wav
if __name__ == '__main__':
sr = 22050
n_fft = 512
win_length = 400
hop_length = 80
duration = 2 # sec
wav = read_wav( "H:\\cs230\\wav_x\\1_1.wav", sr, duration )
spec, _ = wav2spec(wav, n_fft, win_length, hop_length, False)
converted_wav = spec2wav(spec, n_fft, win_length, hop_length, 600)
write_wav(converted_wav, sr, 'a.wav')
plt.pcolormesh(spec)
plt.ylabel('Frequency')
plt.xlabel('Time')
plt.savefig("a.png")
