Rereading our early Icelandic speech recognition papers
Writing two journal papers on conversational speech recognition sent me back to our earlier work, and to the co-authors who led it.
Writing
I have spent the last few weeks writing. Most academics feel they should write more, so this has been a good and guilt-free period. We submitted one journal paper a couple of weeks ago and are well on the way to a full draft of a second. Both are on automatic speech recognition for conversational speech, and I hope to describe them in more detail once they are accepted and published.
Writing is also a chance to take stock of the field and re-evaluate some of your own older work. Doing so, I want to acknowledge two co-authors who led that work.
Open speech recognition for Icelandic
Anna Björk Nikulásdóttir led the first open-source speech recognition work in our lab and in Iceland. She is first author of Open ASR for Icelandic: Resources and a Baseline System, presented at the Language Resources and Evaluation Conference in 2018. Rereading it reminded me how much later work rests on it. The paper is thorough, covering the resources, a complete recognition system and its evaluation, and the system compared well with the state of the art at the time.
Speech recognition for parliament
Inga Rún Helgadóttir developed a more specialised system for Althingi, the Icelandic parliament, described in Building an ASR Corpus Using Althingi’s Parliamentary Speeches at Interspeech 2017 and The Althingi ASR System at Interspeech 2019. Aligning a large amount of parliamentary audio with its text, normalising that text, choosing the modelling approach and evaluating the result were each substantial pieces of work. The outcome was a system that, I believe, is still used at Althingi to produce transcripts of parliamentary speeches.
Anna and Inga co-authored each other’s papers, and Matthías Pétursson, Róbert Kjaran, Michal Borský and Judy Y. Fong were co-authors on one or more of them, as was I.