Establishing a Baseline of Romanian Speech-to-Text Models

With the increasing usage of Natural Language Processing to facilitate the interactions between humans and machines, automatic speech recognition systems have become increasingly popular as a result of their utility in a wide range of applications. In this paper we explore well-known open-source speech-to-text engines, namely CMUSphinx, DeepSpeech, and Kaldi, to build a baseline of models to transcribe Romanian speech. These engines employ various underlying methods from hidden Markov models to deep neural networks that also integrate language models, thus providing a solid baseline for comparison. Unfortunately, Romanian is still a low-resource language and six datasets of various qualities were merged to obtain 104 hours of speech. To further increase the size of the gathered corpora, our experiments consider data augmentation techniques, specifically SpecAugment, applied on the most promising model. Besides using existing corpora, we publicly release a dataset of 11.5 hours generated from Governmental transcripts. The best performing model is obtained using the Kaldi architecture, considers a hybrid structure with a Deep Neural Network, and achieves a WER of 3.10% on the test partition.