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):
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 do_convert(predictor, input_name, logdir2):
convert_s = datetime.datetime.now()
# Load input audio
input_audio, _ = librosa.load(input_name, sr=hp.default.sr, dtype=np.float64)
# Extract F0 from input audio first
input_f0, t_table = pw.dio(input_audio, hp.default.sr)
input_f0 = pw.stonemask(input_audio, input_f0, t_table, hp.default.sr)
# Get MFCC, Spectral Envelope, and Aperiodicity
mfcc = _get_mfcc(input_audio, hp.default.n_fft, hp.default.win_length, hp.default.hop_length)
mfcc = np.expand_dims(mfcc, axis=0)
input_ap = pw.d4c(input_audio, input_f0, t_table, hp.default.sr, fft_size=hp.default.n_fft)
input_sp_en = _get_spectral_envelope(preemphasis(input_audio, coeff=hp.default.preemphasis), hp.default.n_fft)
plt.imsave('./converted/debug/input_sp_en_original.png', input_sp_en, cmap='binary')
input_sp_en = np.expand_dims(input_sp_en, axis=0)
# Convert Spectral Envelope
output_sp_en, ppgs = convert_spectral_envelope(predictor, mfcc, input_sp_en)
output_sp_en = np.squeeze(output_sp_en.astype(np.float64), axis=0)
preproc_s = datetime.datetime.now()
# Denormalization
output_sp_en = denormalize_db(output_sp_en, hp.default.max_db, hp.default.min_db)
# Db to amp
output_sp_en = librosa.db_to_amplitude(output_sp_en)
# Emphasize the magnitude
output_sp_en = np.power(output_sp_en, hp.convert.emphasis_magnitude)
preproc_e = datetime.datetime.now()
preproc_t = preproc_e - preproc_s
print("Pre-Processing time:{}s".format(preproc_t.seconds))
# F0 transformation with WORLD Vocoder
output_f0 = f0_adapt(input_f0, logdir2)
# Synthesize audio and de-emphasize
output_audio = pw.synthesize(output_f0, output_sp_en, input_ap, hp.default.sr)
output_audio = inv_preemphasis(output_audio, coeff=hp.default.preemphasis)
# Saving output_audio to 32-bit Float wav file
output_audio = output_audio.astype(np.float32)
librosa.output.write_wav(path="./converted/"+input_name,y=output_audio,sr=hp.default.sr)
# Saving PPGS data to Grayscale Image and raw binary file
ppgs = np.squeeze(ppgs, axis=0)
plt.imsave('./converted/debug/'+input_name+'.png', ppgs, cmap='binary')
np.save('./converted/debug/'+input_name+'.npy', ppgs)
convert_e = datetime.datetime.now()
convert_time = convert_e - convert_s
print("Total Converting Time:{}s".format(convert_time.seconds))
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 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, mfcc, spec, mel_spec):
print("convert")
pred_s = datetime.datetime.now()
pred_spec, _, ppgs = predictor(mfcc, spec, mel_spec)
pred_e = datetime.datetime.now()
pred_t = pred_e - pred_s
print("Predicting time:{}s".format(pred_t.seconds))
preproc_s = datetime.datetime.now()
# Denormalizatoin
print("denormalize_db")
pred_spec = denormalize_db(pred_spec, hp.default.max_db, hp.default.min_db)
# Db to amp
print("db2amp")
pred_spec = db2amp(pred_spec)
# Emphasize the magnitude
print("emphasize")
pred_spec = np.power(pred_spec, hp.convert.emphasis_magnitude)
preproc_e = datetime.datetime.now()
preproc_t = preproc_e - preproc_s
print("Pre-Processing time:{}s".format(preproc_t.seconds))
audio = []
# Spectrogram to waveform
recon_s = datetime.datetime.now()
print("spec2wav")
audio.append(
spec2wav_lws(pred_spec[0], hp.default.n_fft, hp.default.win_length,
hp.default.hop_length, hp.default.lws_mode))
recon_e = datetime.datetime.now()
recon_t = recon_e - recon_s
print("Converting Spectrogram-to-Wave time:{}s".format(recon_t.seconds))
audio = np.array(audio)
# print('audio.shape : ', audio.shape)
# Apply inverse pre-emphasis
audio = inv_preemphasis(audio, coeff=hp.default.preemphasis)
return audio[0], ppgs
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
