SpeeD's DNN approach to Romanian speech recognition

This paper presents the main improvements brought recently to the large-vocabulary, continuous speech recognition (LVCSR) system for Romanian language developed by the Speech and Dialogue (SpeeD) research laboratory. While the most important improvement consists in the use of DNN-based acoustic models, instead of the classic HMM-GMM approach, several other aspects are discussed in the paper: a significant increase of the speech training corpus, the use of additional algorithms for feature processing, speaker adaptive training, and discriminative training and, finally, the use of lattice rescoring with significantly expanded language models (n-gram models up to order 5, based on vocabularies of up to 200k words). The ASR experiments were performed with several types of acoustic and language models in different configurations on the standard read and conversational speech corpora created by SpeeD in 2014. The results show that the extension of the training speech corpus leads to a relative word error rate (WER) improvement between 15% and 17%, while the use of DNN-based acoustic models instead of HMM-GMM-based acoustic models leads to a relative WER improvement between 18% and 23%, depending on the nature of the evaluation speech corpus (read or conversational, clean or noisy). The best configuration of the LVCSR system was integrated as a live transcription web application available online on SpeeD laboratory's website at https://speed.pub.ro/live-transcriber-2017.