Recent advances in ASR Applied to an Arabic transcription system for Al-Jazeera

/in , , /by
Recent advances in ASR Applied to an Arabic transcription system for Al-Jazeera

Cardinal, Patrick and Ali, Ahmed and Dehak, Najim and Zhang, Yu and Al Hanai, Tuka and Zhang, Yifan and Glass, James and Vogel, Stephan

Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH 2014

Abstract : This paper describes a detailed comparison of several state-of-the-art speech recognition techniques applied to a limited Arabic broadcast news dataset. The different approaches were all trained on 50 hours of transcribed audio from the Al-Jazeera news channel. The best results were obtained using i-vector-based speaker adaptation in a training scenario using the Minimum Phone Error (MPE) criteria combined with sequential Deep Neural Network (DNN) training. We report results for two different types of test data: broadcast news reports, with a best word error rate (WER) of 17.86%, and a broadcast conversations with a best WER of 29.85%. The overall WER on this test set is 25.6%.

You might also like
Handwritten digits segmentation based on structural approach
A multi-neighborhood multi-operator algorithm for the uncapacitated exam proximity problem
Image segmentation in a kernel-induced space
Forest species recognition based on ensembles of classifiers
GNN-DES: A new end-to-end dynamic ensemble selection method based on multi-label graph neural network
Reinforcement learning: adaptive agents for discovery of policies of action
Native language detection using the i-vector framework
Combining classifiers in the ROC-space for off-line signature verification