W266_Project

W266 Project - Automatic Speech Recognition using Transducers

Base model

Implemented in se2seq.ipynb file

Model Parameters

• LSTM RNN with hidden layer size 1024
• Both Encoder and Decoder use 2 layer cascaded LSTMs
• Dropout of 0.2 and teacher forcing ratio of 90%
• ADAM optimizer with exponential decay of learning rate

Transducer model

Implemented in transducer_asr.ipynb file

Model Parameters

• GRU RNN with hidden layer size 1024
• The Encoder uses 3 layer cascaded GRU and decoder uses one layer GRU
• Dropout of 0.1
• ADAM optimizer with exponential decay of learning rate

Transducer model with speech input

Implemented in asr_train.ipynb file

• Speech data samples at 16KHz, Hanning window of 20msec and 10mse overlap
• 64 MFCC coefficient computer using 512 DFT
• Computationally very intensive - hence did not train the model completely