Skip to Main Content
Table 4:
Speech Spectrum Prediction Test on LSTM, GRU, EURNN, GORU, uRNN, and oRNN.
Model Number of Units MSE(validation) MSE(test) 
LSTM 50 51.0 50.7 
GRU 60 52.1 52.4 
EURNN(Jing et al., 2016) 128 51.8 51.9 
oRNN 128 46.2 46.9 
uRNN 108 --- --- 
GORU (EURNN FFT-style) 64 45.5 45.7 
GORU (Householder) 64 40.9 43.0 
GORU (with ReLU) 64 45.8 47.4 
GORU (with tanh) 64 59.7 59.6 
GORU (without reset gate) 64 45.9 46.9 
GORU (without update gate) 64 46.3 47.9 
Model Number of Units MSE(validation) MSE(test) 
LSTM 50 51.0 50.7 
GRU 60 52.1 52.4 
EURNN(Jing et al., 2016) 128 51.8 51.9 
oRNN 128 46.2 46.9 
uRNN 108 --- --- 
GORU (EURNN FFT-style) 64 45.5 45.7 
GORU (Householder) 64 40.9 43.0 
GORU (with ReLU) 64 45.8 47.4 
GORU (with tanh) 64 59.7 59.6 
GORU (without reset gate) 64 45.9 46.9 
GORU (without update gate) 64 46.3 47.9 

Note: The hidden size of each model is set to match the total number of parameters. uRNN failed to converge in this task. GORU significantly outperforms all other RNN models.

Close Modal

or Create an Account

Close Modal
Close Modal