tf.sg_arg_def(frac=(1.0, "test fraction ratio to whole data set. The default is 1.0(=whole set)")) # # hyper parameters # # batch size batch_size = 16 # # inputs # # corpus input tensor ( with QueueRunner ) data = SpeechCorpus(batch_size=batch_size, set_name=tf.sg_arg().set) # mfcc feature of audio x = data.mfcc # target sentence label y = data.label # sequence length except zero-padding seq_len = tf.not_equal(x.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1) # # Testing Graph # # encode audio feature logit = get_logit(x, voca_size=voca_size)
"test fraction ratio to whole data set. The default is 1.0(=whole set)")) # # hyper parameters # # batch size batch_size = 16 # # inputs # # corpus input tensor ( with QueueRunner ) print 'building corpus' data = SpeechCorpus(batch_size=batch_size, set_name=tf.sg_arg().set) # mfcc feature of audio x = data.mfcc # target sentence label y = data.label filenames_t = data.filenames # sequence length except zero-padding seq_len = tf.not_equal(x.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1) # # Testing Graph #
seq_len,
merge_repeated=False)
# to dense tensor
y = tf.sparse_to_dense(decoded[0].indices, decoded[0].dense_shape,
decoded[0].values) + 1
#
# regcognize wave file
#
# command line argument for input wave file path
tf.sg_arg_def(file=('', 'speech wave file to recognize.'))
# load wave file
wav, _ = librosa.load(tf.sg_arg().file, mono=True, sr=16000)
# get mfcc feature
mfcc = np.transpose(np.expand_dims(librosa.feature.mfcc(wav, 16000), axis=0),
[0, 2, 1])
# run network
with tf.Session() as sess:
# init variables
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
# run session
label = sess.run(y, feed_dict={x: mfcc})
#decoded, _ = tf.nn.ctc_beam_search_decoder(logit.sg_transpose(perm=[1, 0, 2]), seq_len, merge_repeated=False)
# to dense tensor
#y = tf.sparse_to_dense(decoded[0].indices, decoded[0].shape, decoded[0].values) + 1
y = tf.nn.softmax(logit)
#
# regcognize wave file
#
# command line argument for input wave file path
tf.sg_arg_def(file=('', 'speech wave file to recognize.'))
# load wave file
wav, sr = librosa.load(tf.sg_arg().file, mono=True)
# get mfcc feature
mfcc = np.transpose(
np.expand_dims(librosa.feature.mfcc(wav, sr, n_mfcc=20), axis=0),
[0, 2, 1])
mfcc = mfcc[:, :100]
# run network
with tf.Session() as sess:
# init variables
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
import data
tf.sg_verbosity(10)
batch_size = 1 # batch size
vocab_size = data.vocab_size
x = tf.placeholder(dtype=tf.sg_floatx, shape=(batch_size, None, 20))
seq_len = tf.not_equal(x.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1)
logit = get_logit(x, voca_size=vocab_size)
decoded, _ = tf.nn.ctc_beam_search_decoder(logit.sg_transpose(perm=[1, 0, 2]), seq_len, merge_repeated=False)
y = tf.sparse_to_dense(decoded[0].indices, decoded[0].dense_shape, decoded[0].values) + 1
tf.sg_arg_def(file=('', 'speech wave file to recognize.'))
wav, _ = librosa.load(tf.sg_arg().file, mono=True, sr=16000)
mfcc = np.transpose(np.expand_dims(librosa.feature.mfcc(wav, 16000), axis=0), [0, 2, 1])
with tf.Session() as sess:
tf.sg_init(sess)
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
label = sess.run(y, feed_dict={x: mfcc})
data.print_index(label)
tf.sg_arg_def(frac=(1.0, "test fraction ratio to whole data set. The default is 1.0(=whole set)")) # # hyper parameters # # batch size batch_size = 16 # # inputs # # corpus input tensor ( with QueueRunner ) data = SpeechCorpus(batch_size=batch_size, set_name=tf.sg_arg().set) # mfcc feature of audio x = data.mfcc # target sentence label y = data.label # sequence length except zero-padding seq_len = tf.not_equal(x.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1) # # Testing Graph # # encode audio feature logit = get_logit(x, voca_size=voca_size)
