RSC: A Romanian Read Speech Corpus for Automatic Speech Recognition

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6606–6612 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6606 RSC: A Romanian Read Speech Corpus for Automatic Speech Recognition Alexandru-Lucian Georgescu, Horia Cucu, Andi Buzo, Corneliu Burileanu Speech and Dialogue Research Laboratory University Politehnica of Bucharest Bucharest, Romania lucian.georgescu@speed.pub.ro, horia.cucu@upb.ro, buzo.andi@gmail.com, corneliu.burileanu@upb.ro Abstract Although many efforts have been made in the last decade to enhance the speech and language resources for Romanian, this language is still considered under-resourced. While for many other languages there are large speech corpora available for research and commercial applications, for Romanian language the largest publicly available corpus to date comprises less than 50 hours of speech. In this context, Speech and Dialogue research group releases Read Speech Corpus (RSC) – a Romanian speech corpus developed in-house, comprising 100 hours of speech recordings from 164 different speakers. The paper describes the development of the corpus and presents baseline automatic speech recognition (ASR) results using state-of-the-art ASR technology: Kaldi speech recognition toolkit. Keywords: speech corpus, automatic speech recognition 1. Introduction Verbal communication plays a crucial role in our lives. The advance of engineering has brought the possibility of com- munication between human and machine. Speech technolo- gies have reached some maturity in recent years and in- tense efforts in the research area led to the emergence of many commercial products. At present, there is a wide range of voice user interfaces, integrated in various applica- tions, such as operating systems, both for personal comput- ers and for smartphones, home and car control systems or other hands-free applications. There are already attempts to create portable translators – pocket-size devices helping people who speak different languages to understand each other in real time. The algorithms involved in all these applications are able to perform both automatic speech recognition and speech synthesis and they are activated based on the spoken hot- word detection mechanism. The very high performances that they achieve are thanks to the use of deep neural net- works together with graphic cards for intense computation. Most of the best algorithms can be found in open source toolkits, along with the data preparation and system train- ing recipes. Although obtaining a speech recognition system seems quite easy, the biggest problem is still represented by the availability of acoustic and linguistic resources. Large amounts of data are required to train such systems and they are not public or simply do not exist for many of the spoken languages. While this problem does not arise for English (for this language there are corpora that contain thousands and tens of thousands hours of speech), Romanian language is still struggling with a shortage of resources that can be used in speech technology systems. The most important speech resources in Romanian are pre- sented in Table 1. We can notice that not all of them are public and the amount of data has the order of hours or tens of hours, in few cases …

Content truncated to 500 words