metadata

dataset_info:
  features:
    - name: speaker_id
      dtype: string
    - name: gender
      dtype: string
    - name: utterance_id
      dtype: string
    - name: language
      dtype: string
    - name: raw_text
      dtype: string
    - name: full_audio_file
      dtype: string
    - name: original_data_split
      dtype: string
    - name: region
      dtype: string
    - name: duration
      dtype: float64
    - name: start
      dtype: float64
    - name: end
      dtype: float64
    - name: utterance_audio_file
      dtype: audio
    - name: standardized_text
      dtype: string
  splits:
    - name: train
      num_bytes: 2078070414.74
      num_examples: 6630
  download_size: 1624762124
  dataset_size: 2078070414.74

Dataset Card for NB Tale, module 1 and 2 (< 15 sec. segments)

Dataset Description

Homepage:
Repository: https://github.com/scribe-project/nodalida_2023_combined_training
Paper:


@inproceedings{
solberg2023improving,
title={Improving Generalization of Norwegian {ASR} with Limited Linguistic Resources},
author={Per Erik Solberg and Pablo Ortiz and Phoebe Parsons and Torbj{\o}rn Svendsen and Giampiero Salvi},
booktitle={The 24rd Nordic Conference on Computational Linguistics},
year={2023}
}

Point of Contact: Per Erik Solberg

Dataset Summary

This is the version of the Bokmål segments of module 1 and 2 of NB Tale used for testing the models in the paper Improving Generalization of Norwegian ASR with Limited Linguistic Resources presented at NoDaLiDa 2023. It only contains segments of a length < 15 sec. This dataset contains both native and non-native speakers. Speakers with region set to foreign were filtered out when analyzing the data in the paper.

Languages

Norwegian Bokmål

Dataset Creation

Source Data

The full version of this dataset is found in the repository of the Norwegian Language Bank

Initial Data Collection and Normalization

The data was retrieved using the Spraakbanken downloader and standardized using the combined dataset standardization scripts. Bokmål segments with a duration < 15 seconds were extracted using this code.

Licensing Information

CC0

Citation Information


@inproceedings{
solberg2023improving,
title={Improving Generalization of Norwegian {ASR} with Limited Linguistic Resources},
author={Per Erik Solberg and Pablo Ortiz and Phoebe Parsons and Torbj{\o}rn Svendsen and Giampiero Salvi},
booktitle={The 24rd Nordic Conference on Computational Linguistics},
year={2023}
}