Deep Learning Methods for Arabic Autoencoder Speech Recognition System for Electro-Larynx Device

Table 2

The best chosen model structure for Arabic speech recognition.

Different specifications for each model
Name of each layer	Autoencoder1	Autoencoder2	Autoencoder3
Encoder layer 1 (120 units) dropout	LSTM (120 units) 0.25 rate	GRU (120 units) 0.25 rate	LSTM (120 units) 0.25 rate
Encoder layer 2	LSTM (60 units)	GRU (60 units)	LSTM (120 units)
Dropout	0.25 rate	0.25 rate	0.25 rate
Encoder layer 3	LSTM (120 units)	GRU (120 units)	LSTM (120 units)
Decoder layer 1	LSTM (120 units)	GRU (120 units)	GRU (120 units)
Decoder layer 2	LSTM (60 units)	GRU (60 units)	GRU (60 units)
Dropout	0.25 rate	0.25 rate	0.25 rate
Flatten layer	—	—	—
Recognize layer
Dense layer	SoftMax (10 neurons)	SoftMax (10 neurons)	SoftMax (10 neurons)