Table IV.

Comparison of Recurrent Neural Network models In Speech Processing

Architecture	Application	Contribution	Limitations
Amodei et al. [159] Gated Recurrent Unit Network	English or Chinese Speech Recognition	Optimized speech recognition using Gated Recurrent Units to achieve near human-level results	Deployment requires GPU server
Weston et al. [134] Memory Network	Answering questions about simple text stories	Integration of long term memory (readable and writable) component within neural network architecture	Questions and input stories are still rather simple
Wu et al. [136] Deep LSTM	Language Translation (e.g. English-to-French)	Multi-layer LSTM with attention mechanism	Challenging translation cases and multisentence input yet to be tested
Karpathy et al. [137] CNN/RNN Fusion	Labeling Images and Image Regions	Hybrid CNN-RNN model to generate natural language descriptions of images	Fixed image size / requires training CNN and RNN models separately