def griffin_lim(mel, hparams=None):
"""
Inverts a mel spectrogram using Griffin-Lim. The mel spectrogram is expected to have been built
with the same parameters present in hparams.py.
"""
hparams = hparams or default_hparams
return audio.inv_melspectrogram(mel, hparams)
## Run a test
print("Testing your configuration with small inputs.")
print("\tTesting the encoder...")
encoder.embed_utterance(np.zeros(encoder.sampling_rate))
embed = np.random.rand(speaker_embedding_size)
embed /= np.linalg.norm(embed)
embeds = [embed, np.zeros(speaker_embedding_size)]
texts = ["你好", "欢迎使用语音克隆工具"]
print("\tTesting the synthesizer... (loading the model will output a lot of text)")
mels = synthesizer.synthesize_spectrograms(texts, embeds)
mel = np.concatenate(mels, axis=1)
no_action = lambda *args: None
generated_wav = audio.inv_melspectrogram(mel, hparams=audio.melgan_hparams)
print("All test passed! You can now synthesize speech.\n\n")
print("Interactive generation loop")
num_generated = 0
args.out_dir.mkdir(exist_ok=True, parents=True)
while True:
try:
# Get the reference audio filepath
message = "Reference voice: enter an audio filepath of a voice to be cloned (mp3, " \
"wav, m4a, flac, ...):\n"
ref = input(message)
in_fpath = Path(ref.replace("\"", "").replace("\'", ""))
if not in_fpath.is_file():
in_fpath = np.random.choice(reference_paths)
print('Reference audio: {}'.format(in_fpath))
