samromur_milljon / README.md
carlosdanielhernandezmena's picture
Typo in BibTex Citation
25ba1ba
metadata
annotations_creators:
  - crowdsourced
language:
  - is
language_creators:
  - crowdsourced
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: Samrómur Milljón, Audio and Transcriptions
size_categories:
  - 1M<n<10M
source_datasets:
  - original
tags:
  - crowd-sourced icelandic
  - samrómur
  - icelandic speech
  - samromur
  - iceland
task_categories:
  - automatic-speech-recognition
task_ids: []

Dataset Card for samromur_milljon

Table of Contents

Dataset Description

Dataset Summary

Samrómur Milljón is the result of the automatic verification of the corpus Samromur Unverified 22.07 which contains 2,159,314 (2233 hours) speech-recordings in Icelandic that are essentially unverified. Samrómur Milljón consists of 1,002,157 of speech recordings (967 hours). The transcripts accompanying these recordings were automatically verified using various ASR systems such as: Wav2Vec2, Whisper, Faster-Whisper and NeMo. If any of the ASR systems get a transcript that perfectly matches the reference transcription, the speech-recording is then cosidered as verified.

Example Usage

Samrómur Milljón is divided in 7 portions. To load a specific split pass its name as a config name:

from datasets import load_dataset
samromur_milljon = load_dataset("language-and-voice-lab/samromur_milljon")

To load a specific split (for example, the split with female speakers between 18 and 49 years old) do:

from datasets import load_dataset
samromur_milljon = load_dataset("language-and-voice-lab/samromur_milljon",split="female_18to49_yrs")

Supported Tasks

automatic-speech-recognition: The dataset can be used to train a model for Automatic Speech Recognition (ASR). The model is presented with an audio file and asked to transcribe the audio file to written text. The most common evaluation metric is the word error rate (WER).

Languages

The audio is in Icelandic. The reading prompts were gathered from a variety of sources, mainly from the Icelandic Gigaword Corpus. The corpus includes text from novels, news, plays, and from a list of location names in Iceland. The prompts also came from the Icelandic Web of Science.

Dataset Structure

Data Instances

{
  'audio_id': '000664-0013663',
  'audio': {
    'path': '/home/carlos/.cache/HuggingFace/datasets/downloads/extracted/5238762551eb35654e659f0bf3f3bf30ba02b40036715337500cba84051fed4e/FEM_18to49_YRS_PART_0001/000664/000664-0013663.flac',
    'array': array([0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 0.0000000e+00,
       3.0517578e-05, 3.0517578e-05], dtype=float32),
    'sampling_rate': 16000
  },
  'speaker_id': '000664',
  'gender': 'female',
  'age': '30-39',
  'duration': 3.7799999713897705,
  'verified_with': 'V+F',
  'normalized_text': 'mattías lækkaðu í hátalaranum'
}

Data Fields

  • audio_id (string) - id of audio segment
  • audio (datasets.Audio) - a dictionary containing the path to the audio, the decoded audio array, and the sampling rate. In non-streaming mode (default), the path points to the locally extracted audio. In streaming mode, the path is the relative path of an audio inside its archive (as files are not downloaded and extracted locally).
  • speaker_id (string) - id of speaker
  • gender (string) - gender of speaker (male or female)
  • age (string) - range of age of the speaker.
  • duration (float32) - duration of the audio file in seconds.
  • verified_with (string) - Set of ASR systems that found a perfect match between the current speech file and the reference transcription : V = wav2vec, N = NeMo, W = Whisper, F = Whisper-Faster
  • normalized_text (string) - normalized audio segment transcription.

Data Splits

Due to its large size, the corpus is divided in 7 splits with the objective that it will be more manageable this way.

  • Female speakers less than 18 years old (n < 18): female_lt_18_yrs

  • Female speakers from 18 to 49 years old (18 <= n <=49): female_18to49_yrs

  • Female speakers greater than 49 years old (n > 49): female_gt_49_yrs

  • Male speakers less than 18 years old (n < 18): male_lt_18_yrs

  • Male speakers from 18 to 49 years old (18 <= n <=49): male_18to49_yrs

  • Male speakers greater than 49 years old (n > 49): male_gt_49_yrs

  • Speakers where age, gender or both are unknown: other

To load an specific portion please see the above section "Example Usage".

Dataset Creation

Curation Rationale

  • The collection process started in October 2019.

  • The aim is to create an open-source speech corpus to enable research and development for Icelandic Language Technology.

  • The corpus comprises of audio recordings and metadata files containing the text sentences read by the participants.

  • At the beginning, some data was manually verified by summer students, but due to the success of the collection process, we ended up with more than 2 million of unverified speech-recordings. In 2023, it was decided to use the distict ASR systems that we already had in Icelandic to automatically verify as much data as possible. The verification process consited in transcribing the recordings using various ASR systems. If one or more of them get a transcription that perfectly matches the reference transcription, the recording is considered as verified.

  • The resulting corpus of more than 1 million recordings is too big to download it all at once. In consequence, it was decided to split it in 7 portions with the hope that it will be more manageable this way.

Source Data

Initial Data Collection and Normalization

Annotations

Annotation process

Prompts were pulled from these corpora if they met the criteria of having only letters which are present in the Icelandic alphabet, and if they are listed in the DIM: Database Icelandic Morphology.

There are also synthesised prompts consisting of a name followed by a question or a demand, in order to simulate a dialogue with a smart-device.

Who are the annotators?

The audio files content was automatically verified using ASR systems such as: Wav2Vec2, Whisper, Faster-Whisper and NeMo. If any of the ASR systems get a transcript that perfectly matches the reference transcription, the speech-recording is then cosidered as verified.

Personal and Sensitive Information

The dataset consists of people who have donated their voice. You agree to not attempt to determine the identity of speakers in this dataset.

Considerations for Using the Data

Social Impact of Dataset

This contribution describes a project of speech data collection, using the web application Samrómur which is built upon Common Voice, Mozilla Foundation's web platform for open-source voice collection. The goal of the project is to build a large-scale speech corpus for Automatic Speech Recognition (ASR) for Icelandic. Samrómur is the largest open speech corpus for Icelandic collected from the public domain.

Discussion of Biases

  • The participants are aged between 4 to 90 years old. The speech files were recorded by a smartphone or the web app.

  • Participants self-reported their age group, gender, and the native language.

  • The corpus contains 1,002,157 from 16,729 speakers, totalling 967 hours and 7 minutes.

  • The number of recordings coming from female speakers are 714,564, totalling 697h22m.

  • The number of recordings coming from male speakers are 282,499, totalling 264h28m.

  • The number of recordings where the gender, the age or both is unknown are 5,094, totalling 5h16m. These recordings belong to the split called "other".

Other Known Limitations

"Samrómur Milljón" by the Language and Voice Laboratory (LVL) from Reykjavik University (RU) is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) License with the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Additional Information

Dataset Curators

The corpus was curated by Carlos Daniel Hernández Mena during 2023 at the Language and Voice Laboratory from Reykjavik University.

Licensing Information

CC-BY-4.0

Citation Information

@misc{menasamromurmilljon2023,
      title={Samrómur Milljón, Audio and Transcriptions}, 
      author={Hernández Mena, Carlos Daniel and Guðnason, Jón},
      publisher={Reykjavík University},
      year={2023},
      url={https://huggingface.co/datasets/language-and-voice-lab/samromur_milljon}
}

Contributions

This project was funded by the Language Technology Programme for Icelandic 2019-2023. The programme, which is managed and coordinated by Almannarómur, is funded by the Icelandic Ministry of Education, Science and Culture.