def test_magspec_to_waveform(self):
x = self.wav_mono_22
self.assertEqual(x.shape, (82432, 1, 1), 'invalid shape')
self.assertEqual(x.dtype, np.float32)
X_mag = spectral.stft(x, 1024, 256, pad_end=False)
self.assertEqual(X_mag.shape, (319, 513, 1), 'invalid shape')
np.random.seed(0)
x_gl0 = spectral.magspec_to_waveform_griffin_lim(X_mag, 1024, 256, ngl=0)
x_gl60 = spectral.magspec_to_waveform_griffin_lim(X_mag, 1024, 256, ngl=60)
x_gl100 = spectral.magspec_to_waveform_griffin_lim(X_mag, 1024, 256, ngl=100)
x_lws = spectral.magspec_to_waveform_lws(X_mag, 1024, 256)
self.assertEqual(x_gl0.shape, (82432, 1, 1), 'invalid shape')
self.assertEqual(x_gl0.dtype, np.float32, 'invalid dtype')
self.assertEqual(x_gl60.shape, (82432, 1, 1), 'invalid shape')
self.assertEqual(x_gl0.dtype, np.float32, 'invalid dtype')
self.assertEqual(x_gl100.shape, (82432, 1, 1), 'invalid shape')
self.assertEqual(x_gl100.dtype, np.float32, 'invalid dtype')
self.assertEqual(x_lws.shape, (82432, 1, 1), 'invalid shape')
self.assertEqual(x_lws.dtype, np.float32, 'invalid dtype')
x_gl0_l1 = np.mean(np.abs(x_gl0 - x))
self.assertAlmostEqual(x_gl0_l1, 0.0232695210048, 8, 'bad l1 after GL0')
x_gl60_l1 = np.mean(np.abs(x_gl60 - x))
self.assertAlmostEqual(x_gl60_l1, 0.0310892466788, 8, 'bad l1 after GL60')
x_gl100_l1 = np.mean(np.abs(x_gl100 - x))
self.assertAlmostEqual(x_gl100_l1, 0.0281844304033, 8, 'bad l1 after GL100')
x_lws_l1 = np.mean(np.abs(x_lws - x))
self.assertAlmostEqual(x_lws_l1, 0.0004236908353, 8, 'bad l1 after LWS')
spec_fn = os.path.splitext(os.path.split(spec_fp)[1])[0]
wave_fn = spec_fn + '.wav'
wave_fp = os.path.join(args.out_dir, wave_fn)
spec = np.load(spec_fp)
if heuristic:
wave = r9y9_melspec_to_waveform(spec)
else:
subseq_len = args.subseq_len
X_mag = tacotron_mel_to_mag(spec[:, :, 0], inv_mel_filterbank)
x_mag_original_length = X_mag.shape[0]
x_mag_target_length = int(
X_mag.shape[0] / subseq_len) * subseq_len + subseq_len
X_mag = np.pad(X_mag,
([0, x_mag_target_length - X_mag.shape[0]], [0, 0]),
'constant')
num_examples = int(x_mag_target_length / subseq_len)
X_mag = np.reshape(X_mag, [num_examples, subseq_len, 513, 1])
gen_mags = []
for n in range(num_examples):
_gen = gen_sess.run([gen_mag_spec],
feed_dict={x_mag_input: X_mag[n:n + 1]})[0]
gen_mags.append(_gen[0])
gen_mag = np.concatenate(gen_mags, axis=0)
gen_mag = gen_mag[0:x_mag_original_length]
wave = magspec_to_waveform_lws(gen_mag.astype('float64'), 1024,
256)
save_as_wav(wave_fp, args.fs, wave)
def main():
parser = ArgumentParser()
parser.add_argument('--input_dir', type=str)
parser.add_argument('--output_dir', type=str)
parser.add_argument('--meta_fp', type=str)
parser.add_argument('--ckpt_fp', type=str)
parser.add_argument('--heuristic', type=str)
parser.add_argument('--n_mels', type=int)
parser.add_argument('--fs', type=int)
parser.add_argument('--subseq_len', type=int)
parser.set_defaults(input_file=None,
output_dir=None,
ckpt_fp=None,
meta_fp=None,
heuristic="lws",
n_mels=80,
fs=22050,
subseq_len=256)
args = parser.parse_args()
if not os.path.isdir(args.output_dir):
os.makedirs(args.output_dir)
gen_graph = tf.Graph()
with gen_graph.as_default():
gan_saver = tf.train.import_meta_graph(args.meta_fp)
gen_sess = tf.Session(graph=gen_graph)
print("Restoring")
gan_saver.restore(gen_sess, args.ckpt_fp)
gen_mag_spec = gen_graph.get_tensor_by_name(
'generator/decoder_1/strided_slice_1:0')
x_mag_input = gen_graph.get_tensor_by_name('ExpandDims_1:0')
su = spectral_util.SpectralUtil(n_mels=args.n_mels, fs=args.fs)
spec_fps = glob.glob(os.path.join(args.input_dir, '*.npy'))
subseq_len = args.subseq_len
start = time.time()
for fidx, fp in enumerate(spec_fps):
_mel_spec = np.load(fp)[:, :, 0]
X_mag = su.tacotron_mel_to_mag(_mel_spec)
x_mag_original_length = X_mag.shape[0]
x_mag_target_length = int(
X_mag.shape[0] / subseq_len) * subseq_len + subseq_len
X_mag = np.pad(X_mag,
([0, x_mag_target_length - X_mag.shape[0]], [0, 0]),
'constant')
num_examples = int(x_mag_target_length / subseq_len)
X_mag = np.reshape(X_mag, [num_examples, subseq_len, 513, 1])
gen_mags = []
heuristic_mags = []
for n in range(num_examples):
_gen, _heur = gen_sess.run([gen_mag_spec, x_mag_input],
feed_dict={x_mag_input: X_mag[n:n + 1]})
_gen = np.clip(_gen, 0, None)
gen_mags.append(_gen[0])
heuristic_mags.append(_heur[0])
gen_mag = np.concatenate(gen_mags, axis=0)
heur_mag = np.concatenate(heuristic_mags, axis=0)
_gen_audio = su.audio_from_mag_spec(gen_mag)
gen_mag = gen_mag[0:x_mag_original_length]
if args.heuristic == 'lws':
_gen_audio = spectral.magspec_to_waveform_lws(
gen_mag.astype('float64'), 1024, 256)
elif args.heuristic == 'gl':
_gen_audio = spectral.magspec_to_waveform_griffin_lim(
gen_mag, 1024, 256)
else:
raise NotImplementedError()
fn = fp.split("/")[-1][:-3] + "wav"
output_file_name = os.path.join(args.output_dir, fn)
print("Writing", fidx, output_file_name)
audioio.save_as_wav(output_file_name, args.fs, _gen_audio)
end = time.time()
print("Execution Time in Seconds", end - start)