Improvements in Serbian speech recognition using sequence-trained deep neural networks

This article presents the recent improvements in Serbian speech recognition that were obtained by using contemporary deep neural networks based on sequence-discriminative training to train robust acoustic models. More specifically, several variants of the new large vocabulary continuous speech recognition (LVCSR) system are described, all based on the lattice-free version of the maximum mutual information (LF-MMI) training criterion. The parameters of the system were varied to achieve best possible word error rate (WER) and character error rate (CER), using the largest speech database for Serbian in existence and the best $n$-gram based language model made for general purposes. In addition to tuning the neural network itself – its layers, complexity, layer splicing and more – other language-specific optimizations were explored, such as the usage of accent-specific vowel phoneme models, and its combination with pitch features to produce the best possible results. Finally, speech database tuning was tested as well – artificial database expansion by modifying speech speed in utterances, as well as volume scaling in an attempt to improve speech variability.
The results suggest that 8-layer deep neural network with moderately sized 625-neuron layers works best in the given environment, without the need for speech database augmentation or volume adjustments, and that pitch features in combination with the introduction of accented vowel models provide the best performance out of all experiments.