MustafaElkhateeb/ASR-Egypt
Domaine:
natural language processingType de record:
project
Créateur:
Mus
Hôte:
This is the repository of Voiceless Voice team for the AIC Competition 2
This is the repository of Voiceless Voice team for the AIC Competition 2
- Speech Preprocessing
- Text Preprocessing
- Working Techniques
- Feature Extraction
- Model Building and training
------------------------------------------------------
Audio Preprocessing
- Remove noise from speech
- Trim audio to remove silence
-------------------------------------------------------
Remove non-Arabic Alphabet
chars_to_ignore = ['>', ']', 'ل', 'ـ', ' ', 1: 'ر', 2: 'l', 3: 'ً', 4: 'ا', 5: 'ب', 6: 'م', 7: 'ك', 8: 'ق', 9: 'ز', 10: 'ص', 11: 'آ', 12: 'ئ', 13: 'غ', 14: 'ى', 15: 'ط', 16: 'ؤ', 17: 'ت', 18: 'ظ', 19: 'ج', 20: 'ذ', 21: 'ش', 22: 'إ', 23: 'خ', 24: 'ن', 25: 'ة', 26: ' ', 27: 'ي', 28: 'ه', 29: 'ء', 30: 'ح', 31: 'و', 32: 'ث', 33: 'ع', 34: 'أ', 35: 'س', 36: 'ف', 37: 'د', 38: 'ض', 39: ' '}
----------------------------------------------------------
Speech-Transformer
Encoder(
(lstm): LSTM(13, 128, batch_first=True, bidirectional=True)
(pBLSTMs): Sequential(
(0): pBLSTM(
(blstm1): LSTM(512, 128, batch_first=True, bidirectional=True)
(lock): LockedDropout()
)
(1): pBLSTM(
(blstm1): LSTM(512, 128, batch_first=True, bidirectional=True)
(lock): LockedDropout()
)
(2): pBLSTM(
(blstm1): LSTM(512, 128, batch_first=True, bidirectional=True)
(lock): LockedDropout()
)
)
(key_network): Linear(in_features=256, out_features=128, bias=True)
(value_network): Linear(in_features=256, out_features=128, bias=True)
)
Bidirectional LSTM:
The Bi-LSTM layer processes the input sequence in both forward and backward directions simultaneously. During the forward pass
Teacher forcing:
Teacher forcing acts like "training wheels." If the model makes a bad prediction, it is put back in place with the true value. Finally, we can make predictions using mixed teacher forcing.