Will the Icelandic Language Technology program and our efforts become role models for other similar languages? Read about RU’s take on the work we are participating in and see us in action (picture taken pre-Covid).
Gæti íslenska máltækniáætlunin orðið leiðarljós annarra lítilla málsvæða? Hérna er umfjöllun HR um starfsemi okkar (Myndir teknar fyrir Covid).
We now have a Facebook page for our spoken Icelandic collection efforts, Samrómur. We’re working on enabling school children to also participate. Follow Samrómur to be told when that platform goes live: https://www.facebook.com/samromur/
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 platformsamromur.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 fromRÚV (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.
We are collecting audio clips to create a new, more robust Icelandic dataset. We plan to use them for automatic speech recognition and text-to-speech applications. This could lead to many applications that you can use in your everyday life, like an Icelandic voice assistant, screenreaders, and searchable audio and video content. The dataset will soon be available directly on the Samromur website for users and developers alike to download and use to creating exciting new technology. But to do this, we need your voices, Icelandic speakers.
You can get started by going to Samromur.is and choosing Tala. You only need to donate 5 recordings per session, but more is always welcome. If you want to do more, you can also evaluate clips made by other users with Hlusta.
We are also organizing data gathering competitions within companies. So, please contact us if you and your company would be interested in holding a competition.
This is a joint effort with Deloitte and Almannaromur. If you want to learn more about this project, visit the Um Verkefnið tab on the Samromur site.
September 2019 signals the end of the LVL automatic speech recognition, ASR, project with Althingi, the Icelandic parliament. To close, the radio station Rás 1 is airing an interview September 8 at 9:30pm (21:30). The interview is conducted by the head of the Althingi speech department, Berglind Steinsdóttir. In the interview, Berglind talks to both Inga Rún, our ASR expert, and Steinunn, an Althingi editor. They discuss both sides of the project: software development and user experiences. This broadcast will hopefully give our Icelandic readers and listeners a deeper understanding of the specifics involved in ASR. Thus, we invite you to tune in this Sunday at 21:30.
Practical information about the radio program is below:
Date: Sunday, September 8th @ 21:30 (re-airing Saturday, September 14th 20:45)
Location: Rás 1 website or the radio station
Duration: 30 minutes
Title: “Háttvirtur þingmaður tekur til máls”
Topic: Sjálfvirknivæðing. Gervigreind. Fjórða iðnbyltingin. Hvað kemur þetta ræðum þingmanna við? Tekinn hefur verið í notkun talgreinir sem skrifar upp ræður þingmanna og í þættinum er rætt við Ingu Rún Helgadóttur eðlisfræðing sem hefur tekið þátt í að þróa hann og Steinunni Haraldsdóttur íslenskufræðing sem hefur notað talgreininn.
Language of Interview: Icelandic
Interviewees: Inga Rún Helgadóttir, ASR developer
Steinunn Haraldsdóttir, icelandic specialist who uses the ASR
Interviewer: Berglind Steinsdóttir
Supervisor: Ásdís Emilsdóttir Petersen.
During the interview go to https://ruv.is/ras1 Click on Í BEINNI. Select Rás 1 and press the play button.
This has been a great summer for LVL. We have many conference acceptances: 8 papers, 3 conferences. It will also be a busy autumn, as all the conferences are in September.
Our first is Recent Advances in Natural Language Processing, a very competitive NLP conference. This year it is in Varna, Bulgaria. Steinþór, Örvar and Hrafn’s paper, “Augmenting a BiLSTM tagger with a Morphological Lexicon and a Lexical Category Identification Step” and Hrafn’s paper “A Wide-Coverage Context-Free Grammar for Icelandic and an Accompanying Parsing System” will both be presented.
Next is Interspeech in Graz, Austria. We have four papers:
Yu-Ren Chien – “F0 Variability Measures Based on Glottal Closure Instants”
Inga Helgadóttir – “The Althingi ASR System”
Anna Rúnarsdóttir – “Lattice re-scoring during manual editing for automatic error correction of ASR transcripts”
Anna Nikulásdóttir – “Bootstraping a Text Normalization System for an Inflected Language. Numbers as a Test Case”