Dataset Card for ravnursson_asr

Dataset Summary

The corpus "RAVNURSSON FAROESE SPEECH AND TRANSCRIPTS" (or RAVNURSSON Corpus for short) is a collection of speech recordings with transcriptions intended for Automatic Speech Recognition (ASR) applications in the language that is spoken at the Faroe Islands (Faroese). It was curated at the Reykjavík University (RU) in 2022.

The RAVNURSSON Corpus is an extract of the "Basic Language Resource Kit 1.0" (BLARK 1.0) [1] developed by the Ravnur Project from the Faroe Islands [2]. As a matter of fact, the name RAVNURSSON comes from Ravnur (a tribute to the Ravnur Project) and the suffix "son" which in Icelandic means "son of". Therefore, the name "RAVNURSSON" means "The (Icelandic) son of Ravnur". The double "ss" is just for aesthetics.

The audio was collected by recording speakers reading texts. The participants are aged 15-83, divided into 3 age groups: 15-35, 36-60 and 61+.

The speech files are made of 249 female speakers and 184 male speakers; 433 speakers total. The recordings were made on TASCAM DR-40 Linear PCM audio recorders using the built-in stereo microphones in WAVE 16 bit with a sample rate of 48kHz, but then, downsampled to 16kHz@16bit mono for this corpus.

[1] Simonsen, A., Debess, I. N., Lamhauge, S. S., & Henrichsen, P. J. Creating a basic language resource kit for Faroese. In LREC 2022. 13th International Conference on Language Resources and Evaluation.

[2] Website. The Project Ravnur under the Talutøkni Foundation https://maltokni.fo/en/the-ravnur-project

Example Usage

The RAVNURSSON Corpus is divided in 3 splits: train, validation and test. To load a specific split pass its name as a config name:

from datasets import load_dataset
ravnursson = load_dataset("carlosdanielhernandezmena/ravnursson_asr")

To load an specific split (for example, the validation split) do:

from datasets import load_dataset
ravnursson = load_dataset("carlosdanielhernandezmena/ravnursson_asr",split="validation")

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 Faroese. The reading prompts for the RAVNURSSON Corpus have been generated by expert linguists. The whole corpus was balanced for phonetic and dialectal coverage; Test and Dev subsets are gender-balanced. Tabular computer-searchable information is included as well as written documentation.

Dataset Structure

Data Instances

{
  'audio_id': 'KAM06_151121_0101', 
  'audio': {
    'path': '/home/carlos/.cache/HuggingFace/datasets/downloads/extracted/32b4a757027b72b8d2e25cd9c8be9c7c919cc8d4eb1a9a899e02c11fd6074536/dev/RDATA2/KAM06_151121/KAM06_151121_0101.flac', 
    'array': array([ 0.0010376 , -0.00521851, -0.00393677, ...,  0.00128174,
        0.00076294,  0.00045776], dtype=float32), 
    'sampling_rate': 16000
  }, 
  'speaker_id': 'KAM06_151121', 
  'gender': 'female', 
  'age': '36-60', 
  'duration': 4.863999843597412, 
  'normalized_text': 'endurskin eru týdningarmikil í myrkri', 
  'dialect': 'sandoy'
}

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: Younger (15-35), Middle-aged (36-60) or Elderly (61+).
• duration (float32) - duration of the audio file in seconds.
• normalized_text (string) - normalized audio segment transcription
• dialect (string) - dialect group, for example "Suðuroy" or "Sandoy".

Data Splits

The speech material has been subdivided into portions for training (train), development (evaluation) and testing (test). Lengths of each portion are: train = 100h08m, test = 4h30m, dev (evaluation)=4h30m.

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

The development and test portions have exactly 10 male and 10 female speakers each and both portions have exactly the same size in hours (4.5h each).

Dataset Creation

Curation Rationale

The directory called "speech" contains all the speech files of the corpus. The files in the speech directory are divided in three directories: train, dev and test. The train portion is sub-divided in three types of recordings: RDATA1O, RDATA1OP and RDATA2; this is due to the organization of the recordings in the original BLARK 1.0. There, the recordings are divided in Rdata1 and Rdata2.

One main difference between Rdata1 and Rdata2 is that the reading environment for Rdata2 was controlled by a software called "PushPrompt" which is included in the original BLARK 1.0. Another main difference is that in Rdata1 there are some available transcriptions labeled at the phoneme level. For this reason the audio files in the speech directory of the RAVNURSSON corpus are divided in the folders RDATA1O where "O" is for "Orthographic" and RDATA1OP where "O" is for Orthographic and "P" is for phonetic.

In the case of the dev and test portions, the data come only from Rdata2 which does not have labels at the phonetic level.

It is important to clarify that the RAVNURSSON Corpus only includes transcriptions at the orthographic level.

Source Data

Initial Data Collection and Normalization

The dataset was released with normalized text only at an orthographic level in lower-case. The normalization process was performed by automatically removing punctuation marks and characters that are not present in the Faroese alphabet.

Who are the source language producers?

• The utterances were recorded using a TASCAM DR-40.

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

• Participants are aged between 15 to 83 years.

• The corpus contains 71949 speech files from 433 speakers, totalling 109 hours and 9 minutes.

Annotations

Annotation process

Most of the reading prompts were selected by experts from a Faroese text corpus (news, blogs, Wikipedia etc.) and were edited to fit the format. Reading prompts that are within specific domains (such as Faroese place names, numbers, license plates, telling time etc.) were written by the Ravnur Project. Then, a software tool called PushPrompt were used for reading sessions (voice recordings). PushPromt presents the text items in the reading material to the reader, allowing him/her to manage the session interactively (adjusting the reading tempo, repeating speech productions at wish, inserting short breaks as needed, etc.). When the reading session is completed, a log file (with time stamps for each production) is written as a data table compliant with the TextGrid-format.

Who are the annotators?

The corpus was annotated by the Ravnur Project

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 is the first ASR corpus in Faroese.

Discussion of Biases

As the number of reading prompts was limited, the common denominator in the RAVNURSSON corpus is that one prompt is read by more than one speaker. This is relevant because is a common practice in ASR to create a language model using the prompts that are found in the train portion of the corpus. That is not recommended for the RAVNURSSON Corpus as it counts with many prompts shared by all the portions and that will produce an important bias in the language modeling task.

In this section we present some statistics about the repeated prompts through all the portions of the corpus.

• In the train portion:

  • Total number of prompts = 65616
  • Number of unique prompts = 38646 There are 26970 repeated prompts in the train portion. In other words, 41.1% of the prompts are repeated.

• In the test portion:

  • Total number of prompts = 3002
  • Number of unique prompts = 2887 There are 115 repeated prompts in the test portion. In other words, 3.83% of the prompts are repeated.

• In the dev portion:

  • Total number of prompts = 3331
  • Number of unique prompts = 3302 There are 29 repeated prompts in the dev portion. In other words, 0.87% of the prompts are repeated.

• Considering the corpus as a whole:

  • Total number of prompts = 71949
  • Number of unique prompts = 39945 There are 32004 repeated prompts in the whole corpus. In other words, 44.48% of the prompts are repeated.

NOTICE!: It is also important to clarify that none of the 3 portions of the corpus share speakers.

Other Known Limitations

"RAVNURSSON FAROESE SPEECH AND TRANSCRIPTS" by Carlos Daniel Hernández Mena and Annika Simonsen 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 dataset was collected by Annika Simonsen and curated by Carlos Daniel Hernández Mena.

Licensing Information

CC-BY-4.0

Citation Information

@misc{carlosmenaravnursson2022,
      title={Ravnursson Faroese Speech and Transcripts}, 
      author={Hernandez Mena, Carlos Daniel and Simonsen, Annika},
      year={2022},
      url={http://hdl.handle.net/20.500.12537/276},
}

Contributions

This project was made possible under the umbrella of 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.

Special thanks to Dr. Jón Guðnason, professor at Reykjavík University and head of the Language and Voice Lab (LVL) for providing computational resources.