class SpeechTranscriber:
def __init__(self, model_path):
"""
:param model_path:
"""
assert os.path.exists(model_path), "Cannot find model here {}".format(
model_path)
self.deep_speech_model = DeepSpeech.load_model(model_path)
self.deep_speech_model.eval()
labels = DeepSpeech.get_labels(self.deep_speech_model)
self.audio_conf = DeepSpeech.get_audio_conf(self.deep_speech_model)
self.decoder = GreedyDecoder(labels)
self.parser = SpectrogramParser(self.audio_conf, normalize=True)
def transcribe(self, audio_file):
"""
:param audio_file:
:return:
"""
spect = self.parser.parse_audio(audio_file).contiguous()
spect = spect.view(1, 1, spect.size(0), spect.size(1))
out = self.deep_speech_model(Variable(spect, volatile=True))
out = out.transpose(0, 1) # TxNxH
decoded_output = self.decoder.decode(out.data)
return decoded_output
default='hamming',
help='Window type for spectrogram generation')
parser.add_argument('--cuda',
default=True,
type=bool,
help='Use cuda to train model')
args = parser.parse_args()
if __name__ == '__main__':
package = torch.load(args.model_path)
model = DeepSpeech(rnn_hidden_size=package['hidden_size'],
nb_layers=package['hidden_layers'],
num_classes=package['nout'])
if args.cuda:
model = torch.nn.DataParallel(model).cuda()
model.load_state_dict(package['state_dict'])
audio_conf = dict(sample_rate=args.sample_rate,
window_size=args.window_size,
window_stride=args.window_stride,
window=args.window)
with open(args.labels_path) as label_file:
labels = str(''.join(json.load(label_file)))
decoder = ArgMaxDecoder(labels)
parser = SpectrogramParser(audio_conf, normalize=True)
spect = parser.parse_audio(args.audio_path).contiguous()
spect = spect.view(1, 1, spect.size(0), spect.size(1))
out = model(Variable(spect))
out = out.transpose(0, 1) # TxNxH
decoded_output = decoder.decode(out.data)
print(decoded_output[0])
# while recv_data != b'ok':
# recv_data = connection.recv(1024)
# print('ok waiting')
# print('ok received')
recv_data = connection.recv(1024)
recv_file = open(file_name, 'wb')
while recv_data:
recv_file.write(recv_data)
recv_data = connection.recv(1024)
recv_file.close()
print('download complete')
start = time.time()
#inference
spect = parser.parse_audio(file_name).contiguous()
parsing_time = time.time() - start
spect = spect.view(1, 1, spect.size(0), spect.size(1))
out = model(spect)
inferring_time = time.time() - parsing_time - start
decoded_output, decoded_offsets = decoder.decode(out.data)
decoding_time = time.time() - inferring_time - start
print('time for parsing: %0.4f,\t inferring: %0.4f,\t decoding: %0.4f'%(parsing_time,inferring_time,decoding_time))
print(json.dumps(decode_results(model, decoded_output, decoded_offsets)))
except ConnectionResetError:#connection is broken by a client
pass
finally:
# Clean up the connection