Open ASR for Icelandic: Resources and a Baseline System

Open ASR for Icelandic: Resources and a Baseline System Anna Björk Nikulasdóttir, Inga Rún Helgadóttir, Matthías Pétursson, Jón Guðnason Reykjavik University Reykjavik Iceland {annabn, ingarun, matthias, jg}@ru.is Abstract Developing language resources is an important task when creating a speech recognition system for a less-resourced language. In this paper we describe available language resources and their preparation for use in a large vocabulary speech recognition (LVSR) system for Icelandic. The content of a speech corpus is analysed and training and test sets compiled, a pronunciation dictionary is extended, and text normalization for language modeling performed. An ASR system based on neural networks is implemented using these resources and tested using different acoustic training sets. Experimental results show a clear increase in word-error-rate (WER) when using smaller training sets, indicating that extension of the speech corpus for training would improve the system. When testing on data with known vocabulary only, the WER is 7.99%, but on an open vocabulary test set the WER is 15.72%. Furthermore, impact of the content of the acoustic training corpus is examined. The current results indicate that an ASR system could profit from carefully selected phonotactical data, however, further experiments are needed to verify this impression. The language resources are available on http://malfong.is and the source code of the project can be found on https://github.com/cadia-lvl/ice-asr/tree/master/ice-kaldi. Keywords: language resources, automatic speech recognition, Icelandic 1. Introduction With advances in automatic speech recognition (ASR) tech- nology and access to open source ASR tools like Kaldi1 (Povey et al., 2011), the threshold for the development of ASR systems has been lowered substantially. The core ASR algorithms, including state-of-the-art neural network architectures, are already implemented, leaving the devel- opment of language resources as the main task. The development of sufficient language resources, how- ever, poses a challenge for less-resourced languages. Two specialized resources are needed for automatic speech recognition: a speech corpus with matching units of speech and text, and a pronunciation dictionary. Additionally, a large text corpus is needed for the development of a lan- guage model. In this paper we describe the development of language re- sources for an Icelandic large vocabulary speech recogni- tion system (LVSR), and the setup of a baseline system in Kaldi. The first speech recognizer for Icelandic was Hjal, devel- oped in 2003 as an isolated word recognizer (Rögnvalds- son, 2003). The Hjal project was successful as a pilot project, but has not been maintained. Currently, three ASR systems are being developed for Icelandic using Kaldi: an ASR system for parliamentary speeches (Helgadóttir et al., 2017), a system for automating transcriptions of radiology dictations2, and the LVSR project described in this paper. Google already supports Icelandic in its speech recognition applications3. We inspect available language resources for Icelandic and prepare them for use in a speech recognition system. Af- ter defining the characteristics of the resources and testing an initial setup of an LSVR system, we make suggestions for further development of the resources. In particular we 1http://kaldi-asr.org/ 2http://laeknaromur.is/en.html 3see e.g. https://cloud.google.com/speech/ address three questions: …

Content truncated to 500 words