With the advent of weakly supervised ASR systems like Whisper, it is possible to train ASR systems on non-verbatim transcriptions. This paper describes an effort to create a large Norwegian dataset for weakly supervised ASR from parliamentary recordings. Audio from Stortinget, the Norwegian parliament, is segmented and transcribed with an existing ASR system. An algorithm retrieves transcripts of these segments from Stortinget’s official proceedings using the Levenshtein edit distance between the ASR output and the proceedings text. In that way, a dataset of more than 5000 hours of transcribed speech is produced with limited human effort. Since parliamentary data is public domain, the dataset can be shared freely without any restrictions.

In this paper, we present several baselines for automatic speech recognition (ASR) models for the two official written languages in Norway: Bokmål and Nynorsk. We compare the performance of models of varying sizes and pre-training approaches on multiple Norwegian speech datasets. Additionally, we measure the performance of these models against previous state-of-the-art ASR models, as well as on out-of-domain datasets. We improve the state of the art on the Norwegian Parliamentary Speech Corpus (NPSC) from a word error rate (WER) of 17.10% to 7.60%, with models achieving 5.81% for Bokmål and 11.54% for Nynorsk. We also discuss the challenges and potential solutions for further improving ASR models for Norwegian.

Norwegian has been one of many languages lacking sufficient available text to train quality language models. In an attempt to bridge this gap, we introduce the Norwegian Colossal Corpus (NCC), which comprises 49GB of clean Norwegian textual data containing over 7B words. The NCC is composed of different and varied sources, ranging from books and newspapers to government documents and public reports, showcasing the various uses of the Norwegian language in society. The corpus contains mainly Norwegian Bokmål and Norwegian Nynorsk. Each document in the corpus is tagged with metadata that enables the creation of sub-corpora for specific needs. Its structure makes it easy to combine with large web archives that for licensing reasons could not be distributed together with the NCC. By releasing this corpus openly to the public, we hope to foster the creation of both better Norwegian language models and multilingual language models with support for Norwegian.

In this work, we show the process of building a large-scale training set from digital and digitized collections at a national library. The resulting Bidirectional Encoder Representations from Transformers (BERT)-based language model for Norwegian outperforms multilingual BERT (mBERT) models in several token and sequence classification tasks for both Norwegian Bokmål and Norwegian Nynorsk. Our model also improves the mBERT performance for other languages present in the corpus such as English, Swedish, and Danish. For languages not included in the corpus, the weights degrade moderately while keeping strong multilingual properties. Therefore, we show that building high-quality models within a memory institution using somewhat noisy optical character recognition (OCR) content is feasible, and we hope to pave the way for other memory institutions to follow.