def featurize(self, audio_clip, overwrite=False):
""" For a given audio clip, calculate the log of its Fourier Transform
Params:
audio_clip(str): Path to the audio clip
"""
return spectrogram_from_file(
audio_clip, step=self.step, window=self.window,
max_freq=self.max_freq, overwrite=overwrite)
def featurize(self, audio_clip, overwrite=False):
""" For a given audio clip, calculate the log of its Fourier Transform
Params:
audio_clip(str): Path to the audio clip
"""
return spectrogram_from_file(
audio_clip, step=self.step, window=self.window,
max_freq=self.max_freq, overwrite=overwrite)
def featurize(self,
audio_clip,
overwrite=False,
save_feature_as_csvfile=False,
noise_percent=0.4,
seq_length=-1):
""" For a given audio clip, calculate the log of its Fourier Transform
Params:
audio_clip(str): Path to the audio clip
"""
return spectrogram_from_file(
audio_clip,
step=self.step,
window=self.window,
max_freq=self.max_freq,
overwrite=overwrite,
save_feature_as_csvfile=save_feature_as_csvfile,
noise_percent=noise_percent,
seq_length=seq_length)
def getTrans(self, wav_file):
res = spectrogram_from_file(wav_file, noise_percent=0)
buck = bisect.bisect_left(self.buckets, len(res))
bucket_key = 1600
res = self.datagen.normalize(res)
d = np.zeros((self.batch_size, bucket_key, res.shape[1]))
d[0, :res.shape[0], :] = res
st = time.time()
# model_loaded.forward(data_batch, is_train=False)
probs = self.forward(mx.nd.array(d))
from stt_metric import ctc_greedy_decode
res = ctc_greedy_decode(probs, self.labelUtil.byList)
self.log.info("forward cost %.2f, %s" % (time.time() - st, res))
st = time.time()
# model_loaded.update_metric(self.eval_metric, data_batch.label)
self.log.info("upate metric cost %.2f" % (time.time() - st))
# print("my res is:")
# print(eval_metric.placeholder)
return self.eval_metric.placeholder
def getTrans(self, wav_file):
res = spectrogram_from_file(wav_file, noise_percent=0)
buck = bisect.bisect_left(self.buckets, len(res))
bucket_key = self.buckets[buck]
res = self.datagen.normalize(res)
d = np.zeros((self.batch_size, bucket_key, res.shape[1]))
d[0, :res.shape[0], :] = res
init_state_arrays = [mx.nd.zeros(x[1]) for x in self.init_states]
model_loaded = self.model
provide_data = [
('data', (self.batch_size, bucket_key, self.width * self.height))
] + self.init_states
data_batch = mx.io.DataBatch([mx.nd.array(d)] + init_state_arrays,
label=None,
bucket_key=bucket_key,
provide_data=provide_data,
provide_label=None)
st = time.time()
model_loaded.forward(data_batch, is_train=False)
probs = model_loaded.get_outputs()[0].asnumpy()
log.info("forward cost %.3f" % (time.time() - st))
st = time.time()
res = ctc_greedy_decode(probs, self.labelUtil.byList)
log.info("greedy decode cost %.3f, result is:\n%s" %
(time.time() - st, res))
beam_size = 5
from stt_metric import ctc_beam_decode
st = time.time()
results = ctc_beam_decode(scorer=self.scorer,
beam_size=beam_size,
vocab=self.labelUtil.byList,
probs=probs)
log.info("beam decode cost %.3f, result is:\n%s" %
(time.time() - st, "\n".join(results)))
return "greedy:\n" + res + "\nbeam:\n" + "\n".join(results)