Discussion
The results indicate that the ensembled model Bidirectional LSTM-RNN
model scored the best results on the Word Error rate, accuracy, and Mean
Edit Distance (MED). Parameter optimization of the experiment setup
influenced the better performance as highlighted by (Apeksha Shewalkar,
2019). This research adopted the random search hyperparameter
optimization method. While previous research has focused on performance
based on different architectures, these results demonstrate that with
parameter optimization good values can be achieved within acceptable
running time days. In this research the number of hidden layer nodes
architecture was limited to 1000 nodes this can be in the future
increased to see the impact on the performance of each model.
Conclusion
This section has evaluated Gated Recurrent Unit (GRU), Simple Recurrent
Neural Network (RNN), Long Short-Term Memory (LSTM), LSTM with a bias,
BLSTM, and BLSTM-RNN model and compared the performance using
LibriSpeech data set. The performance evaluation measure used was Word
Error Rate, Mean Distance Edit, and accuracy as well as the training
loss versus the validation loss as per the dataset. The results show
that the BLSTM-RNN model performed best as compared to the other models.
In future work more tests can be conducted on the same models on
different architectures and the parameters too can be modified to see
how the models will perform.