def test_voice_conversion():
print(" > Run voice conversion inference using YourTTS model.")
model_name = "tts_models/multilingual/multi-dataset/your_tts"
language_id = "en"
speaker_wav = os.path.join(get_tests_data_path(), "ljspeech", "wavs",
"LJ001-0001.wav")
reference_wav = os.path.join(get_tests_data_path(), "ljspeech", "wavs",
"LJ001-0032.wav")
output_path = os.path.join(get_tests_output_path(), "output.wav")
run_cli(
f"tts --model_name {model_name}"
f" --out_path {output_path} --speaker_wav {speaker_wav} --reference_wav {reference_wav} --language_idx {language_id} "
)
def test_d_vector_inference(self):
args = VitsArgs(
spec_segment_size=10,
num_chars=32,
use_d_vector_file=True,
d_vector_dim=256,
d_vector_file=os.path.join(get_tests_data_path(),
"dummy_speakers.json"),
)
config = VitsConfig(model_args=args)
model = Vits.init_from_config(config, verbose=False).to(device)
model.eval()
# batch size = 1
input_dummy = torch.randint(0, 24, (1, 128)).long().to(device)
d_vectors = torch.randn(1, 256).to(device)
outputs = model.inference(input_dummy,
aux_input={"d_vectors": d_vectors})
self._check_inference_outputs(config, outputs, input_dummy)
# batch size = 2
input_dummy, input_lengths, *_ = self._create_inputs(config)
d_vectors = torch.randn(2, 256).to(device)
outputs = model.inference(input_dummy,
aux_input={
"x_lengths": input_lengths,
"d_vectors": d_vectors
})
self._check_inference_outputs(config,
outputs,
input_dummy,
batch_size=2)
def _test_forward_with_d_vector(self, batch_size):
input_dummy, input_lengths, mel_spec, mel_lengths, speaker_ids = self._create_inputs(
batch_size)
d_vector = torch.rand(batch_size, 256).to(device)
# create model
config = GlowTTSConfig(
num_chars=32,
use_d_vector_file=True,
d_vector_dim=256,
d_vector_file=os.path.join(get_tests_data_path(),
"dummy_speakers.json"),
)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
model.train()
print(" > Num parameters for GlowTTS model:%s" %
(count_parameters(model)))
# inference encoder and decoder with MAS
y = model.forward(input_dummy, input_lengths, mel_spec, mel_lengths,
{"d_vectors": d_vector})
self.assertEqual(y["z"].shape, mel_spec.shape)
self.assertEqual(y["logdet"].shape, torch.Size([batch_size]))
self.assertEqual(y["y_mean"].shape, mel_spec.shape)
self.assertEqual(y["y_log_scale"].shape, mel_spec.shape)
self.assertEqual(y["alignments"].shape,
mel_spec.shape[:2] + (input_dummy.shape[1], ))
self.assertEqual(y["durations_log"].shape, input_dummy.shape + (1, ))
self.assertEqual(y["total_durations_log"].shape,
input_dummy.shape + (1, ))
def test_init_multispeaker(self):
config = GlowTTSConfig(num_chars=32)
model = GlowTTS(config)
# speaker embedding with default speaker_embedding_dim
config.use_speaker_embedding = True
config.num_speakers = 5
config.d_vector_dim = None
model.init_multispeaker(config)
self.assertEqual(model.c_in_channels, model.hidden_channels_enc)
# use external speaker embeddings with speaker_embedding_dim = 301
config = GlowTTSConfig(num_chars=32)
config.use_d_vector_file = True
config.d_vector_dim = 301
model = GlowTTS(config)
model.init_multispeaker(config)
self.assertEqual(model.c_in_channels, 301)
# use speaker embedddings by the provided speaker_manager
config = GlowTTSConfig(num_chars=32)
config.use_speaker_embedding = True
config.speakers_file = os.path.join(get_tests_data_path(), "ljspeech",
"speakers.json")
speaker_manager = SpeakerManager.init_from_config(config)
model = GlowTTS(config)
model.speaker_manager = speaker_manager
model.init_multispeaker(config)
self.assertEqual(model.c_in_channels, model.hidden_channels_enc)
self.assertEqual(model.num_speakers, speaker_manager.num_speakers)
# use external speaker embeddings by the provided speaker_manager
config = GlowTTSConfig(num_chars=32)
config.use_d_vector_file = True
config.d_vector_dim = 256
config.d_vector_file = os.path.join(get_tests_data_path(),
"dummy_speakers.json")
speaker_manager = SpeakerManager.init_from_config(config)
model = GlowTTS(config)
model.speaker_manager = speaker_manager
model.init_multispeaker(config)
self.assertEqual(model.c_in_channels, speaker_manager.embedding_dim)
self.assertEqual(model.num_speakers, speaker_manager.num_speakers)
def test_init_from_config(self):
config = GlowTTSConfig(num_chars=32)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
config = GlowTTSConfig(num_chars=32, num_speakers=2)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
self.assertTrue(model.num_speakers == 2)
self.assertTrue(not hasattr(model, "emb_g"))
config = GlowTTSConfig(num_chars=32,
num_speakers=2,
use_speaker_embedding=True)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
self.assertTrue(model.num_speakers == 2)
self.assertTrue(hasattr(model, "emb_g"))
config = GlowTTSConfig(
num_chars=32,
num_speakers=2,
use_speaker_embedding=True,
speakers_file=os.path.join(get_tests_data_path(), "ljspeech",
"speakers.json"),
)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
self.assertTrue(model.num_speakers == 10)
self.assertTrue(hasattr(model, "emb_g"))
config = GlowTTSConfig(
num_chars=32,
use_d_vector_file=True,
d_vector_dim=256,
d_vector_file=os.path.join(get_tests_data_path(),
"dummy_speakers.json"),
)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
self.assertTrue(model.num_speakers == 1)
self.assertTrue(not hasattr(model, "emb_g"))
self.assertTrue(model.c_in_channels == config.d_vector_dim)
def test_init_from_config(self):
config = VitsConfig(model_args=VitsArgs(num_chars=32))
model = Vits.init_from_config(config, verbose=False).to(device)
config = VitsConfig(model_args=VitsArgs(num_chars=32, num_speakers=2))
model = Vits.init_from_config(config, verbose=False).to(device)
self.assertTrue(not hasattr(model, "emb_g"))
config = VitsConfig(model_args=VitsArgs(
num_chars=32, num_speakers=2, use_speaker_embedding=True))
model = Vits.init_from_config(config, verbose=False).to(device)
self.assertEqual(model.num_speakers, 2)
self.assertTrue(hasattr(model, "emb_g"))
config = VitsConfig(model_args=VitsArgs(
num_chars=32,
num_speakers=2,
use_speaker_embedding=True,
speakers_file=os.path.join(get_tests_data_path(), "ljspeech",
"speakers.json"),
))
model = Vits.init_from_config(config, verbose=False).to(device)
self.assertEqual(model.num_speakers, 10)
self.assertTrue(hasattr(model, "emb_g"))
config = VitsConfig(model_args=VitsArgs(
num_chars=32,
use_d_vector_file=True,
d_vector_dim=256,
d_vector_file=os.path.join(get_tests_data_path(),
"dummy_speakers.json"),
))
model = Vits.init_from_config(config, verbose=False).to(device)
self.assertTrue(model.num_speakers == 1)
self.assertTrue(not hasattr(model, "emb_g"))
self.assertTrue(model.embedded_speaker_dim == config.d_vector_dim)
def _test_inference_with_d_vector(self, batch_size):
input_dummy, input_lengths, mel_spec, mel_lengths, speaker_ids = self._create_inputs(
batch_size)
d_vector = torch.rand(batch_size, 256).to(device)
config = GlowTTSConfig(
num_chars=32,
use_d_vector_file=True,
d_vector_dim=256,
d_vector_file=os.path.join(get_tests_data_path(),
"dummy_speakers.json"),
)
model = GlowTTS.init_from_config(config, verbose=False).to(device)
model.eval()
outputs = model.inference(input_dummy, {
"x_lengths": input_lengths,
"d_vectors": d_vector
})
self._assert_inference_outputs(outputs, input_dummy, mel_spec)
def test_d_vector_forward(self):
batch_size = 2
args = VitsArgs(
spec_segment_size=10,
num_chars=32,
use_d_vector_file=True,
d_vector_dim=256,
d_vector_file=os.path.join(get_tests_data_path(),
"dummy_speakers.json"),
)
config = VitsConfig(model_args=args)
model = Vits.init_from_config(config, verbose=False).to(device)
model.train()
input_dummy, input_lengths, _, spec, spec_lengths, waveform = self._create_inputs(
config, batch_size=batch_size)
d_vectors = torch.randn(batch_size, 256).to(device)
output_dict = model.forward(input_dummy,
input_lengths,
spec,
spec_lengths,
waveform,
aux_input={"d_vectors": d_vectors})
self._check_forward_outputs(config, output_dict)
from TTS.tts.datasets import TTSDataset, load_tts_samples
from TTS.tts.utils.text.tokenizer import TTSTokenizer
from TTS.utils.audio import AudioProcessor
# pylint: disable=unused-variable
OUTPATH = os.path.join(get_tests_output_path(), "loader_tests/")
os.makedirs(OUTPATH, exist_ok=True)
# create a dummy config for testing data loaders.
c = BaseTTSConfig(text_cleaner="english_cleaners",
num_loader_workers=0,
batch_size=2,
use_noise_augment=False)
c.r = 5
c.data_path = os.path.join(get_tests_data_path(), "ljspeech/")
ok_ljspeech = os.path.exists(c.data_path)
dataset_config = BaseDatasetConfig(
name="ljspeech_test", # ljspeech_test to multi-speaker
meta_file_train="metadata.csv",
meta_file_val=None,
path=c.data_path,
language="en",
)
DATA_EXIST = True
if not os.path.exists(c.data_path):
DATA_EXIST = False
print(" > Dynamic data loader test: {}".format(DATA_EXIST))