Language and Voice Lab

Last week we celebrated achieving the first milestone in the Language Technology for Icelandic project with a cake!

After a lot of hard work the past few months we achieved the first milestone in Automatic Speech Recognition (ASR), Text-to-Speech (TTS) and Machine Translation (MT).

In ASR, the focus has mostly been on data creating and gathering. 55,000 utterances have been collected (donated by adults) via the crowd-sourcing platform samromur.is (based on Common Voice) with plans to reach 100.000 utterances for the next milestone. The process is being extended to include younger voices in collaboration with schools and authorities. Today we started working with Öldutúnsskóli in Hafnarfjörður. The goal is to reach 80.000 young voice utterances for the next mileston. Additionally, data has been gathered from RÚV (audio, video and subtitles) and CreditInfo (transcriptions). Along with data gathering, the team is also developing tools to post-process Icelandic ASR text for better readability.

In TTS, we successfully created a voice recording client (LOBE) and three reading scripts in order to collect high quality speech and corresponding text data. The reading scripts were created from Risamálheild and seek to maximize diphone coverage. So far 20 hours have been collected from two speakers, male and female. The aim is to finish collecting 20 hours from each speaker early this year. From the collected data two TTS prototypes have been created in Ossian, which extends the Merlin back-end. The current prototypes are quite naive but we have integrated a grapheme-to-phoneme model for the Icelandic language into the prototypes.

In MT, we successfully created a phrase-based statistical machine translation system using the open source tool Moses. Our collaborators at Miðeind created neural machine translation systems based on BiLSTMs and Transformers. The models were trained on the newly available English-Icelandic parallel corpus, ParIce. The systems were then evaluated w.r.t. training time, throughput and BLEU score. The code and 
systems are freely available but are still under development for milestone two. In milestone two we will continue to develop the systems further and adjust them to specific needs of the Icelandic language.