Dataset Datasheet: NST – Nordic Language Technology Norwegian Speech Corpus

1. Motivation

The NST dataset was created to support the development of automatic speech recognition (ASR) and speech-to-text dictation systems in Norwegian. The dataset provides high-quality, annotated audio recordings for linguistic research and ASR system training. Originally developed by Nordisk Språkteknologi Holding AS (NST), the dataset was later acquired and preserved by a consortium of Norwegian institutions, including University of Oslo, University of Bergen, and the National Library of Norway.

2. Composition

Each example in the dataset includes:

• Audio: A 16kHz sampled .mp3 file of spoken Norwegian.
• Text: The transcribed content of the speech.
• Metadata: Rich speaker and recording metadata, such as age, sex, region of birth/youth, recording session info, and technical parameters.

The full dataset contains:

• 9 train shards and 3 test shards for Norwegian (no).
• Variants for close, distant, and both microphone positions.
• Thousands of annotated utterances.

What are the data types?

• Audio data (MP3)
• Structured metadata (JSON)
• Text transcription (Norwegian language)

No direct personal identifiers are included, but demographic metadata such as age, sex, and region is present.

3. Collection Process

The recordings were originally made by NST using multiple speakers under controlled environments with both close and distant microphone setups.

Speakers belong to varying demographics:

• Ages: Varies widely (some errors or missing values may exist)
• Sex: Male and female
• Regional background: Multiple Norwegian dialect regions

4. Preprocessing and Labeling

What preprocessing was done?

• WAV files converted to MP3.
• Non-standard or corrupted metadata (e.g., ages with “ramslie”) is coerced into a usable format (e.g., defaulted to 0).
• Fields such as t0, t1, t2, and number_of_recordings are cast to integers or defaulted.

What labeling is provided?

• Transcription (text field) in Norwegian
• Microphone position (close, distant)
• Linguistic and demographic metadata

5. Uses

What are the intended uses?

• Training and evaluating ASR systems in Norwegian.
• Linguistic research on Norwegian dialects and pronunciation.
• Voice-based human-computer interaction studies.

What are the out-of-scope uses?

• Applications requiring real-time conversational data or informal speech.
• Use in sensitive profiling without additional privacy controls.

6. Distribution

The dataset is distributed under the Apache 2.0 License.

• Hugging Face Dataset Hub: https://huggingface.co/datasets/NbAiLab/NST
• Språkbanken (Norwegian National Library): https://www.nb.no/sprakbanken/en/resource-catalogue/oai-nb-no-sbr-54/

7. Maintenance

The dataset is hosted by NbAiLab on Hugging Face and curated by the National Library of Norway. Users are encouraged to provide feedback or contribute improved versions by contacting: sprakbanken@nb.no