Hugging Face's logo

Datasets:
davidilag
/
FPSC

Tasks:
Automatic Speech Recognition
Modalities:
Audio
Text
Formats:
soundfolder
Languages:
Faroese
Size:
10K - 100K
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1
FPSC / README.md
davidilag's picture
davidilag
Update README.md
189a87b verified
preview code
|
raw
history blame contribute delete
8.15 kB
metadata 
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
language:
  - fo
pretty_name: FPSC
size_categories:
  - 10K<n<100K

FPSC — Faroese Parliament Speech Corpus

FPSC is a large-scale Faroese parliamentary speech corpus constructed from publicly available recordings from Løgtingið, the Parliament of the Faroe Islands.

The dataset contains approximately:

  • 1,600 hours of speech
  • 89,000+ parliamentary speeches
  • 368 parliamentary meetings
  • 75 unique speakers
  • speaker demographic metadata
  • dialect metadata
  • machine-generated weak transcripts
  • ROVER voting metadata
  • audio aligned at speech level

The corpus was created as part of the paper:

FPSC: A Sustainable Pipeline for Building a Faroese Parliamentary Speech Corpus
Dávid í Lág, Barbara Scalvini, Carlos Mena, Jón Guðnason
LREC 2026

The dataset represents the first large-scale corpus of natural spoken Faroese and is intended for:

  • Automatic Speech Recognition (ASR)
  • Low-resource speech technology
  • Parliamentary speech analysis
  • Sociolinguistic research
  • Dialect research
  • Weakly supervised ASR training
  • Continual pretraining
  • Multilingual transfer learning

Dataset Structure

The dataset follows the Hugging Face Audio dataset format and contains one row per parliamentary speech segment.

Each entry contains:

  • segmented WAV audio
  • machine-generated transcript
  • parliamentary metadata
  • speaker metadata
  • dialect metadata
  • ROVER voting information
  • ASR ensemble metadata

Features

Audio and Speech Fields

Field Description
audio Hugging Face Audio object containing the speech segment
audio_id Unique WAV filename for the speech segment
duration Audio duration in seconds
length Original segment length in seconds
audio_format Audio format (WAV)
sampling_rate Audio sample rate (16 kHz mono)

Transcript Fields

Field Description
text Final machine-generated transcript selected through ROVER voting
normalized_text Normalized version of the transcript for ASR training
winner_text Raw winning transcript from the ROVER voting process
language Spoken language of the segment (Faroese)

Speaker Metadata

Field Description
speaker_id Internal speaker identifier
mp_id Parliament member identifier
name Speaker name
gender Speaker gender
age Speaker age
age_group Speaker age group
date_of_birth Speaker date of birth
city Speaker home city
dialect Dialect region
political_party_affiliation Political party affiliation
mp_url URL to parliament member profile

Parliamentary Metadata

Field Description
id Unique speech segment ID
meeting_id Parliamentary meeting identifier
url URL to the original parliamentary meeting
date Meeting date
time Speech start time
second Start offset in seconds within the meeting
topic Parliamentary agenda topic
contribution_type Type of contribution (speech, remark, chair, etc.)
location Recording location

ROVER Voting Metadata

The final transcripts were generated using an ensemble of four Faroese-adapted ASR systems combined through ROVER voting.

Field Description
rover_vote_type Type of ROVER voting decision
confidence Confidence score assigned by the voting system
winner_model_id Full Hugging Face model ID of the winning ASR model
winner_model_short Short name of the winning ASR model
costs Full ROVER voting cost dictionary stored as JSON
cost_wav2vec2_fo_cpt Voting cost for Wav2Vec2-FO-CPT
cost_whisper_fo Voting cost for Whisper-FO
cost_wav2vec2_fo Voting cost for Wav2Vec2-FO
cost_whisper_no_is_fo Voting cost for Whisper-NO/IS/FO

ASR Models Used for Weak Transcription

The corpus transcripts were generated using four Faroese-adapted ASR systems:

Model Description
Wav2Vec2-FO-CPT Continually pretrained Wav2Vec2 XLS-R model adapted on Faroese parliamentary speech
Wav2Vec2-FO Faroese fine-tuned Wav2Vec2 XLS-R model
Whisper-FO Whisper Large model fine-tuned on Faroese speech
Whisper-NO/IS/FO Multilingual Whisper model fine-tuned on Norwegian, Icelandic, and Faroese

The final transcript for each segment was selected using weighted ROVER voting based on ASR model performance.

Transcript Quality Notice

The transcripts in FPSC are automatically generated and should be treated as weakly supervised labels, not manually verified gold-standard transcriptions.

Although the corpus was generated using multiple ASR systems and ROVER consensus voting, transcription errors remain present, especially for:

  • overlapping speech
  • dialect variation
  • named entities
  • interruptions
  • spontaneous parliamentary speech

The dataset is therefore most suitable for:

  • weakly supervised ASR training
  • continual pretraining
  • large-scale speech modeling
  • speech representation learning
  • sociolinguistic analysis

Loading the Dataset 

from datasets import load_dataset, Audio

ds = load_dataset("davidilag/FPSC")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

print(ds)
print(ds["train"][0])

Citation

If you use FPSC in your research, please cite:

@inproceedings{iLag2026FPSC,
  author = {Dávid í Lág and Barbara Scalvini and Carlos Mena and Jón Guðnason},
  title = {{FPSC}: A Sustainable Pipeline for Building a Faroese Parliamentary Speech Corpus},
  booktitle = {Proceedings of the Language Resources and Evaluation Conference (LREC 2026)},
  year = {2026},
  address = {Palma de Mallorca, Spain},
  publisher = {European Language Resources Association (ELRA)},
  institution = {University of the Faroe Islands},
  keywords = {Faroese, Parliamentary Speech, Automatic Speech Recognition, Weakly-Supervised Transcription, Whisper, Wav2Vec2},
  abstract = {We present FPSC, a 1,600-hour Faroese parliamentary speech corpus comprising approximately 89,000 speeches enriched with detailed speaker and linguistic metadata. The corpus was constructed using a sustainable ASR-assisted pipeline combining speech segmentation, multiple Faroese-adapted ASR systems, and ROVER-based consensus voting for weakly supervised transcription. FPSC represents the first large-scale corpus of natural spoken Faroese and provides an open resource for future research in automatic speech recognition and low-resource language technology.}
}

Repository and Scripts

Processing scripts and pipeline implementation:

Original Data Source

The original parliamentary recordings and metadata were collected from the official website of Løgtingið, the Parliament of the Faroe Islands:

https://www.logting.fo/mal/yvirlit/gerdabokur/

The website provides publicly accessible:

  • parliamentary meeting recordings
  • agendas and meeting protocols
  • speaker order and timestamps
  • meeting dates and metadata
  • parliament member information

FPSC was constructed by automatically downloading, processing, segmenting, and transcribing these publicly available parliamentary sessions into a structured speech corpus suitable for ASR and language technology research.

License

This dataset is released under the CC BY 4.0 license.

The recordings originate from publicly available parliamentary broadcasts from Løgtingið, the Parliament of the Faroe Islands.

Contact

Dávid í Lág
University of the Faroe Islands
MSc. Computer Science, Ph.D. student in Computer Science (2024-2028)
Research area: Automatic Speech Recognition for Low-Resource Languages