Abstract:
Classifier combination is a way to combine the outputs of multiple distinct recognition
systems so as to improve the accuracy of the combined system. The training
data is used to train multiple instances of neural network based recognition
systems (BLSTM). Next the test data is transcribed using each of the trained models.
A post processing technique (ROVER) is used to combine the multiple transcriptions
obtained for each test sequence into a best transcription. This approach yields a higher
transcription accuracy when compared to those obtained with the individual recognition
systems used.
