def main():
if len(sys.argv) >= 2:
f = open(sys.argv[1], 'rb')
else:
f = frontend.default_reader()
for msg in frontend.Frontend(f):
print('---------------------------------')
sys.stdout.write(str(msg))
def main(**settings):
"""
This function returns a Pyramid WSGI application.
"""
listen()
# Load dragon configuration
conf.load("jarvis")
# Setup logging
#logging.config.dictConfig(jarvis.LOGGING)
# Start the frontend greenlets and create the WSGI application
_frontend = frontend.Frontend()
_frontend.start(start_wsgi_server=True, forever=True)
def Text2Speech(self, text, req_id):
time_start = time.time()
# 1. front end
idseq = frontend.Frontend(self.phone2id, self.idim, text)
mat = torch.LongTensor(idseq).view(-1).to(device)
time_frontend = time.time()
# 2. acoustic model inference
c, _, _ = self.model.inference(mat, self.inference_args)
features = c.cpu().detach().numpy()
logging.info("{} Pass Acoustic Model".format(__file__))
time_acoustic = time.time()
req_dir = os.path.join(self.tmp_dir, req_id)
os.makedirs(req_dir, exist_ok=True)
# 3. vocoder
batch_feats = Split(self.overlap, features, self.num_chunk_frame)
logging.info("{} Split {} done".format(__file__, len(batch_feats)))
executor = ProcessPoolExecutor(max_workers=len(batch_feats))
futures = []
num_frame = 0
for i, s1, s2, a1, a2, sub_feats in batch_feats:
num_frame += a2 - a1
futures.append(
executor.submit(
partial(run_vocoder, self.overlap, self.num_chunk_frame,
self.tmp_dir, i, s1, s2, a1, a2, sub_feats)))
batch_pcm = [future.result() for future in futures]
pcm = np.zeros((num_frame, 160), dtype="int16")
index = 0
for one_pcm in batch_pcm:
i, sub_pcm = one_pcm
#print(i, sub_pcm.shape)
start = index
num = sub_pcm.shape[0]
#print(start, num)
pcm[start:start + num] = sub_pcm
index += num
pcm = np.reshape(pcm, -1)[:num_frame * 160]
wav_path = os.path.join(req_dir, req_id + "-wave.wav")
with wave.open(wav_path, 'wb') as fp:
# 1 channel, 16 bits, 16000 sample rate
fp.setparams((1, 2, 16000, 0, 'NONE', 'NONE'))
fp.writeframes(pcm)
audio_dur = num_frame * 0.01
logging.info("{} Wave {:0.3f}s {}".format(__file__, audio_dur,
wav_path))
if self.debug_mode:
# numpy type feature
np.save(os.path.join(req_dir, req_id + "-feats.npy"), features)
# plot
plt.figure()
plt.matshow(np.flip(features.T))
plt.savefig(os.path.join(req_dir, req_id + "-demo.png"),
format="png")
plt.close()
time_vocoder = time.time()
time_count_frontend = time_frontend - time_start
time_count_acoustic = time_acoustic - time_frontend
time_count_vocoder = time_vocoder - time_acoustic
logging.info(
"{} Acoustic:{:0.3f}s Vocoder:{:0.3f}s A+V: {:0.3f}s".format(
__file__, time_count_acoustic, time_count_vocoder,
time_vocoder - time_start))
return wav_path
def main():
native = NinjaNativeFrontend()
for msg in frontend.Frontend():
native.handle(msg)
def restart(self, obj):
call(["halt", "--reboot"])
def quit(self, widget, *event):
gtk.main_quit()
def main():
app = GUI()
app.window.show()
gtk.main()
if __name__ == "__main__":
rontend = None
if 'DISPLAY' not in os.environ:
rontend = frontend.Frontend()
rontend.set_lang()
rontend.startx()
rontend.init_gtk()
rontend.start_wm()
rontend.merge_xres()
try:
import gtk
except ImportError:
sys.exit("pygtk not found.")
sys.exit(main())
def Text2Speech(self, text, req_id):
time_start = time.time()
# 1. front end
idseq = frontend.Frontend(self.phone2id, self.idim, text)
mat = torch.LongTensor(idseq).view(-1).to(device)
time_frontend = time.time()
# 2. acoustic model inference
c, _, _ = self.model.inference(mat, self.inference_args)
features = c.cpu().detach().numpy()
logging.info("{} Pass Acoustic Model".format(__file__))
time_acoustic = time.time()
req_dir = os.path.join(self.tmp_dir, req_id)
os.makedirs(req_dir, exist_ok=True)
# 3. vocoder
f32_batch_path = self.SaveNPartAsF32(self.overlap, features,
self.num_chunk_frame, req_dir,
req_id)
logging.info("{} Save F32 Done".format(__file__))
executor = ProcessPoolExecutor(max_workers=len(f32_batch_path))
futures = []
for sub_req_id, sub_f32_path, s1, s2 in f32_batch_path:
futures.append(
executor.submit(
partial(ShellRunLPCNet, self.overlap, self.num_chunk_frame,
self.vocoder_path, self.tmp_dir, sub_f32_path,
req_id, sub_req_id, s1, s2)))
wav_batch_path = [future.result() for future in futures]
wav_path = os.path.join(req_dir, req_id + "-wave.wav")
CombineAudio(wav_batch_path, wav_path)
logging.info("{} Wave {}".format(__file__, wav_path))
if self.debug_mode:
# numpy type feature
np.save(os.path.join(req_dir, req_id + "-feats.npy"), features)
# plot
plt.figure()
plt.matshow(np.flip(features.T))
plt.savefig(os.path.join(req_dir, req_id + "-demo.png"),
format="png")
plt.close()
time_vocoder = time.time()
time_count_frontend = time_frontend - time_start
time_count_acoustic = time_acoustic - time_frontend
time_count_vocoder = time_vocoder - time_acoustic
logging.info(
"{} Acoustic:{:0.3f}s Vocoder:{:0.3f}s A+V: {:0.3f}s".format(
__file__, time_count_acoustic, time_count_vocoder,
time_vocoder - time_start))
return wav_path