def synthesize1(filename, bpm=80, speaker_id=14, outname="sample.wav"):
tacotron, waveglow, denoiser = model
data = get_data_from_musicxml(filename, bpm)
sampling_rate = 22050
frequency_scaling = 0.4
n_seconds = 90
audio_stereo = np.zeros((sampling_rate * n_seconds, 2), dtype=np.float32)
data_v = list(data.values())[0]
rhythm = data_v['rhythm'].cpu()
pitch_contour = data_v['pitch_contour'].cpu()
text_encoded = data_v['text_encoded'].cpu()
speaker_id = torch.LongTensor([speaker_id]).cpu()
with torch.no_grad():
some_number_i_dont_know_what_is_this = 0 # Seems to be a number from 0 to 10
mel_outputs, mel_outputs_postnet, gate_outputs, alignments_transfer = tacotron.inference_noattention(
(text_encoded, some_number_i_dont_know_what_is_this, speaker_id,
pitch_contour * frequency_scaling, rhythm))
audio = denoiser(waveglow.infer(mel_outputs_postnet, sigma=0.66),
0.01)[0, 0]
audio = audio.cpu().numpy()
pan = 0
audio = panner(audio, pan)
audio_stereo[:audio.shape[0]] += audio
write(outname, sampling_rate, audio)
def singing_voice_v2():
# Singing Voice from Music Score
data = get_data_from_musicxml('data/sinsy/csongdb_f00002_000_en.musicxml', 132, convert_stress=True)
panning = {'Soprano': [-60, -30], 'Alto': [-40, -10], 'Tenor': [30, 60], 'Bass': [10, 40]}
n_speakers_per_part = 4
frequency_scaling = 0.4
n_seconds = 90
audio_stereo = np.zeros((hparams.sampling_rate * n_seconds, 2), dtype=np.float32)
for i, (part, v) in enumerate(data.items()):
rhythm = data[part]['rhythm'].cuda()
pitch_contour = data[part]['pitch_contour'].cuda()
text_encoded = data[part]['text_encoded'].cuda()
for k in range(n_speakers_per_part):
pan = k
# pan = np.random.randint(panning[part][0], panning[part][1])
if any(x in part.lower() for x in ('soprano', 'alto', 'female')):
speaker_id = torch.LongTensor([next(female_speakers)]).cuda()
else:
speaker_id = torch.LongTensor([next(male_speakers)]).cuda()
print("{} MellotronID {} pan {}".format(part, speaker_id.item(), pan))
with torch.no_grad():
mel_outputs, mel_outputs_postnet, gate_outputs, alignments_transfer = mellotron.inference_noattention(
(text_encoded, mel, speaker_id, pitch_contour * frequency_scaling, rhythm))
plot_mel_f0_alignment(mel_outputs_postnet.data.cpu().numpy()[0],
mel_outputs_postnet.data.cpu().numpy()[0],
pitch_contour.data.cpu().numpy()[0, 0],
rhythm.data.cpu().numpy()[:, 0].T)
plt.show()
out_mel = mel_outputs_postnet.data.cpu().numpy()[0]
t0 = time.time()
# wav = aukit.inv_mel_spectrogram()
out_wav = infer_waveform_melgan(out_mel)
print(time.time() - t0)
aukit.save_wav(out_wav, "logs/musicxml_melgan_{}.wav".format(time.strftime("%Y%m%d-%H%M%S")), sr=22050)
aukit.play_audio(out_wav, sr=22050)
t0 = time.time()
audio = denoiser(waveglow.infer(mel_outputs_postnet, sigma=0.8), 0.01)[0, 0]
audio = audio.cpu().numpy()
audio = panner(audio, pan)
print(time.time() - t0)
audio_stereo[:audio.shape[0]] += audio
write("logs/{} {}.wav".format(part, speaker_id.item()), hparams.sampling_rate, audio)
out_wav = audio
aukit.play_audio(out_wav, sr=22050)
with open(GLconfig) as f:
GLconfig = yaml.load(f, Loader=yaml.Loader)
spec0 = mel_outputs_postnet.cpu().detach().numpy()[0,:,:]
wavGL = mel_to_wav(spec0.T,GLconfig)
# GL-write audio
filename = 'iteration-gl-'+str(iteration)+timeStr+'.wav'
sf.write(filename, wavGL,16000)
# mel save
from feats.TTS_Feat import np_save
mel_filename = 'iteration-mel-'+str(iteration)+timeStr+'.npy'
np_save(mel_filename,spec0)
# waveglow
# audio_stereo = np.zeros((hparams.sampling_rate*n_seconds, 2), dtype=np.float32)
if 1:
data = get_data_from_musicxml('data/haendel_hallelujah.musicxml', 132, convert_stress=True)
# print('data=',data)
# print('data.shape=',data.shape)
part = 'Soprano'
audio = denoiser(waveglow.infer(mel_outputs_postnet, sigma=0.8), 0.01)[0, 0]
audio = audio.cpu().numpy()
# panning = {'Soprano': [-60, -30], 'Alto': [-40, -10], 'Tenor': [30, 60], 'Bass': [10, 40]}
# pan = np.random.randint(panning[part][0], panning[part][1])
# audio = panner(audio, pan)
# audio_stereo[:audio.shape[0]] += audio
# audio_stereo = audio_stereo / np.max(np.abs(audio_stereo))
filename = 'iteration-waveglow-'+str(iteration)+timeStr+'.wav'
sf.write(filename, audio, 16000)
# plot
t0 = np.linspace(0,len(audio)-1,len(audio))/hparams.sampling_rate
