python parse.py listfile name
will parse all the feats and phonemes file in listfile, align them, replace phonemes by their ID from phntable and dump the array of aligned feats/phoneme index in file name.pkl
Pickle file contains:
- feature np.array - shape (#nframes,123), float32
- phoneme idx np.array - shape (#nframe), int
- index>phone mapping
- phone>index mapping
- phone>group mapping
Note: phone q was missing in phngroup - new file commited in this repository
Numbers of file in each directory:
train/dr1: 181train/dr2: 363train/dr3: 374train/dr4: 320train/dr5: 326train/dr6: 165train/dr7: 375train/dr8: 107
head -200 feat_train.list > feat_train_small.list
python parse.py feat_train_small.list train_small
tail -100 feat_train.list > feat_dev.list
python parse.py feat_dev.list dev
- Check configuration in
config.py. - launch:
python train.py train_small.pkl dev.pkl
Output for small file above:
read train data (train_small.pkl)
* 200 sequences
* 61 phonemes
read dev data (dev.pkl)
* 100 sequences
configure phone recognition model
Epoch 1 out of 20
10/10 [==========================>...] - ETA: 1s - train loss: 3.9356 - dev accuracy 15.78 - PER 71.33
- new best score!
Epoch 2 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 3.3228 - dev accuracy 20.44 - PER 67.05
- new best score!
Epoch 3 out of 20
10/10 [==========================>...] - ETA: 1s - train loss: 3.0407 - dev accuracy 23.82 - PER 64.05
- new best score!
Epoch 4 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.8391 - dev accuracy 25.93 - PER 62.04
- new best score!
Epoch 5 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.6935 - dev accuracy 27.68 - PER 60.23
- new best score!
Epoch 6 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.6003 - dev accuracy 29.63 - PER 58.53
- new best score!
Epoch 7 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.4916 - dev accuracy 30.41 - PER 58.68
Epoch 8 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.4084 - dev accuracy 31.52 - PER 56.79
- new best score!
Epoch 9 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.3308 - dev accuracy 32.59 - PER 56.43
- new best score!
Epoch 10 out of 20
10/10 [==========================>...] - ETA: 1s - train loss: 2.2734 - dev accuracy 34.09 - PER 55.15
- new best score!
Epoch 11 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.2211 - dev accuracy 34.71 - PER 54.25
- new best score!
Epoch 12 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.1822 - dev accuracy 35.20 - PER 54.00
- new best score!
Epoch 13 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.1416 - dev accuracy 36.14 - PER 53.28
- new best score!
Epoch 14 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.1182 - dev accuracy 36.70 - PER 52.83
- new best score!
Epoch 15 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.0896 - dev accuracy 37.28 - PER 52.46
- new best score!
Epoch 16 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.0749 - dev accuracy 37.45 - PER 52.48
Epoch 17 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.0573 - dev accuracy 37.26 - PER 52.62
Epoch 18 out of 20
10/10 [==========================>...] - ETA: 2s - train loss: 2.0353 - dev accuracy 37.27 - PER 52.53
- early stopping 3 epochs without improvement
The model is:
- a bidirectional LSTM built using
tf.nn.bidirectional_dynamic_rnn- input is the sequence of frames - dropout layer for the training
- a logit layer - generating nphones label
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent