def infer(PROD, user_settings, text, output, fastpitch, vocoder, speaker_i, pace=1.0, pitch_data=None, logger=None, old_sequence=None):
print(f'Inferring: "{text}" ({len(text)})')
sigma_infer = 0.9
stft_hop_length = 256
sampling_rate = 22050
denoising_strength = 0.01
sequence = text_to_sequence(text, ['english_cleaners'])
cleaned_text = sequence_to_text(sequence)
text = torch.LongTensor(sequence)
text = pad_sequence([text], batch_first=True).to(fastpitch.device)
with torch.no_grad():
if old_sequence is not None:
old_sequence = text_to_sequence(old_sequence, ['english_cleaners'])
old_sequence = torch.LongTensor(old_sequence)
old_sequence = pad_sequence([old_sequence], batch_first=True).to(fastpitch.device)
mel, mel_lens, dur_pred, pitch_pred = fastpitch.infer_advanced(text, speaker_i=speaker_i, pace=pace, pitch_data=pitch_data, old_sequence=old_sequence)
if "waveglow" in vocoder:
init_waveglow(user_settings["use_gpu"], fastpitch, vocoder, logger=logger)
audios = fastpitch.waveglow.infer(mel, sigma=sigma_infer)
audios = fastpitch.denoiser(audios.float(), strength=denoising_strength).squeeze(1)
for i, audio in enumerate(audios):
audio = audio[:mel_lens[i].item() * stft_hop_length]
audio = audio/torch.max(torch.abs(audio))
write(output, sampling_rate, audio.cpu().numpy())
del audios
else:
init_hifigan(PROD, fastpitch, user_settings["use_gpu"], vocoder)
y_g_hat = fastpitch.hifi_gan(mel)
audio = y_g_hat.squeeze()
audio = audio * 32768.0
# audio = audio * 2.3026 # This brings it to the same volume, but makes it clip in places
audio = audio.cpu().numpy().astype('int16')
write(output, sampling_rate, audio)
del mel, mel_lens
[pitch, durations] = [pitch_pred.cpu().detach().numpy()[0], dur_pred.cpu().detach().numpy()[0]]
pitch_durations_text = ",".join([str(v) for v in pitch])+"\n"+",".join([str(v) for v in durations])
del pitch_pred, dur_pred, text
return pitch_durations_text +"\n"+cleaned_text
def infer(user_settings,
text,
output,
fastpitch,
hifi_gan,
speaker_i,
pace=1.0,
pitch_data=None):
print(f'Inferring: "{text}" ({len(text)})')
sigma_infer = 0.9
stft_hop_length = 256
sampling_rate = 22050
denoising_strength = 0.01
sequence = text_to_sequence(text, ['english_cleaners'])
cleaned_text = sequence_to_text(sequence)
text = torch.LongTensor(sequence)
text = pad_sequence([text], batch_first=True).to(fastpitch.device)
with torch.no_grad():
mel, mel_lens, dur_pred, pitch_pred = fastpitch.infer_advanced(
text, speaker_i=speaker_i, pace=pace, pitch_data=pitch_data)
if hifi_gan:
y_g_hat = fastpitch.hifi_gan(mel)
audio = y_g_hat.squeeze()
audio = audio * 32768.0
audio = audio.cpu().numpy().astype('int16')
write(output, sampling_rate, audio)
else:
init_waveglow(user_settings["use_gpu"], fastpitch)
audios = fastpitch.waveglow.infer(mel, sigma=sigma_infer)
audios = fastpitch.denoiser(audios.float(),
strength=denoising_strength).squeeze(1)
for i, audio in enumerate(audios):
audio = audio[:mel_lens[i].item() * stft_hop_length]
audio = audio / torch.max(torch.abs(audio))
write(output, sampling_rate, audio.cpu().numpy())
[pitch, durations] = [
pitch_pred.cpu().detach().numpy()[0],
dur_pred.cpu().detach().numpy()[0]
]
pitch_durations_text = ",".join([str(v) for v in pitch]) + "\n" + ",".join(
[str(v) for v in durations])
return pitch_durations_text + "\n" + cleaned_text
def infer_batch(PROD,
user_settings,
linesBatch,
fastpitch,
vocoder,
speaker_i,
logger=None,
old_sequence=None):
print(f'Inferring batch of {len(linesBatch)} lines')
sigma_infer = 0.9
stft_hop_length = 256
sampling_rate = 22050
denoising_strength = 0.01
text_sequences = []
for record in linesBatch:
text = record[0]
sequence = text_to_sequence(text, ['english_cleaners'])
text = torch.LongTensor(sequence)
text_sequences.append(text)
text_sequences = pad_sequence(text_sequences,
batch_first=True).to(fastpitch.device)
with torch.no_grad():
pace = torch.tensor([record[3] for record in linesBatch
]).unsqueeze(1).to(fastpitch.device)
pitch_data = None # Maybe in the future
mel, mel_lens, dur_pred, pitch_pred = fastpitch.infer_advanced(
logger,
text_sequences,
speaker_i=speaker_i,
pace=pace,
pitch_data=pitch_data,
old_sequence=None)
if "waveglow" in vocoder:
init_waveglow(user_settings["use_gpu"],
fastpitch,
vocoder,
logger=logger)
audios = fastpitch.waveglow.infer(mel, sigma=sigma_infer)
audios = fastpitch.denoiser(audios.float(),
strength=denoising_strength).squeeze(1)
for i, audio in enumerate(audios):
audio = audio[:mel_lens[i].item() * stft_hop_length]
audio = audio / torch.max(torch.abs(audio))
output = linesBatch[i][4]
write(output, sampling_rate, audio.cpu().numpy())
del audios
else:
init_hifigan(PROD, fastpitch, user_settings["use_gpu"], vocoder)
y_g_hat = fastpitch.hifi_gan(mel)
audios = y_g_hat.view((y_g_hat.shape[0], y_g_hat.shape[2]))
# audio = audio * 2.3026 # This brings it to the same volume, but makes it clip in places
for i, audio in enumerate(audios):
audio = audio[:mel_lens[i].item() * stft_hop_length]
audio = audio.cpu().numpy()
audio = audio * 32768.0
audio = audio.astype('int16')
output = linesBatch[i][4]
write(output, sampling_rate, audio)
del mel, mel_lens
return ""