projecte-aina/annotated_catalan_common_voice_v17

Catalan Common Voice v17 - metadata annotated version

Dataset Summary

This version of the Catalan sentences of the Common Voice corpus (v17) includes metadata (gender and accent) for 263 speakers annotated by a team of experts.

They correspond to the speakers who have recorded more than 1200 seconds of Speech in v13.

We release these annotations (annotated_accent and annotated_gender fields) along with information about the agreement among the annotators (annotated_accent_agreement and annotated_gender_agreement).

The overall quality of the recordings has also been annotated for these speakers (mean_quality and stdev_quality fields).

Additionally, since for certain users self-declared information appears only in some of their recordings (client_id field), we have created new columns with metadata extracted from other recordings of the same user (propagated_gender and propagated_accents fields).

To facilitate data analysis and classification tasks, we have normalized the information in the free-form accents field to the predefined list used until version 8 of the corpus (propagated_accents_norm). For each user, we have added the fields assigned_accent and assigned_gender. For speakers annotated by the team of experts, this information is derived from their annotations. For other speakers, we have extracted this information from their self-declared data. See annotations for more details.

Supported Tasks and Leaderboards

Automatic Speech Recognition, Gender classification, Accent classification.

Languages

The dataset is in Catalan (ca).

Dataset Structure

Data Instances

{
  'client_id': '69dafb41ddc0ea2785719305fdc5c8d79c4b2829d9f3325bda707dcaa553f95c5fbf4b072970d9004d3e31543fcb2c55e252dc904c4fb5aee2a5e5500df90967', 
  'path': 'common_voice_ca_19909748.mp3', 
  'sentence': 'En el carrer de l'església es troben bona part dels edificis importants de la vila.', 
  'up_votes': 2, 
  'down_votes': 0, 
  'age': 'thirties', 
  'gender': 'male', 
  'accent': 'balear', 
  'variant': '',
  'locale': 'ca', 
  'segment': '',
  'mean quality': '4.0',
  'stdev quality': '0.0',
  'annotated_accent': 'balearic',
  'annotated_accent_agreement': '100.0',
  'annotated_gender': 'male_masculine',
  'annotated_gender_agreement': '100.0',
  'propagated_gender': 'male_masculine',
  'propagated_accents': 'balear',
  'propagated_accents_normalized': 'balearic',
  'assigned_accent': 'balearic',
  'assigned_gender': 'male'
  
}

Data Fields

Most of the data fields come from the original Common Voice corpus:

• client_id(string): An id for which client (voice) made the recording

• path(string): The path to the audio file

• sentence_id(string): An id for the text sentence

• sentence(string): The sentence the user was prompted to speak

• sentence_domain(string): Semantic domain of the sentence

• up_votes(int64): How many upvotes the audio file has received from reviewers

• down_votes(int64): How many downvotes the audio file has received from reviewers

• age(string): Self-reported age of the speaker (e.g. teens, twenties, fifties)

• gender(string): Self-reported gender of the speaker

• accent(string): Self-reported accent of the speaker

• locale(string): The locale of the speaker

• segment(string): Usually an empty field

In this version of the corpus, we have added the following fields:

• annotated_gender(string): Annotated gender by the experts team.

• annotated_gender_agreement (float): Agreement whithin the annotation team about the gender of the speaker.

• annotated_accent (string): Annotated accent by the experts team. The accents considered are: Balearic, Central, Northern, Northwestern, Valencian.

• annotated_accent_agreement (float): Agreement whithin the annotaion team about the accent of the speaker.

• mean quality (float): Mean annotated quality of the speakers' recording.

• stdev quality (float): Deviation in the quality annotation between annotators.

• propagated_gender (string): Self-declared gender as indicated in certain recordings by the user. Speakers that change self-declared gender have been labeled as "other.

• propagated_accents (string): Self-declared accent as indicated in certain recordings by the user. See annotations for more information.

• propagated_accents_normalized (string): Propagated accent, normalized to the closed-options list used until version 7.

• assigned_accent (string): Accent assigned to the speaker.

• assigned_gender (string): Gender assigned to the speaker.

Data Splits

Same splits than in the original Common Voice corpus The reported.tsv file hasn't been annotated because it doesn't contain information about the speakers.

split	sentences	manually annotated	% man annot	self-defined accent	% self-d accent	assig accent	% assig accent	self-defined gender	% sd gender	assig gender	% ag gender
annotated_dev.tsv	16391	0	0.0	2374	14.48	2391	14.59	899	5.48	924	5.64
annotated_train.tsv	1142932	219711	19.22	666672	58.33	836843	73.22	718620	62.88	854209	74.74
annotated_other.tsv	489510	169583	34.64	247831	50.63	423430	86.5	398199	81.35	422862	86.38
annotated_test.tsv	16402	0	0.0	2922	17.81	3024	18.44	755	4.6	785	4.79
annotated_invalidated.tsv	109232	14712	13.47	65596	60.05	74399	68.11	66878	61.23	75802	69.4
annotated_validated.tsv	1817954	368602	20.28	1136712	62.53	1417174	77.95	1241752	68.3	1437806	79.09

Dataset Creation

Curation Rationale

During 2022, a campaign was launched to promote the Common Voice corpus within the Catalan-speaking community, achieving remarkable success. However, not all participants provided their demographic details such as age, gender, and accent. Additionally, some individuals faced difficulty in self-defining their accent using the standard classifications established by specialists.

In order to obtain a balanced corpus with reliable information, we have seen the the necessity of enlisting a group of experts from the University of Barcelona to provide accurate annotations.

We hope that this corpus will be a useful contribution to the development of ASR tasks in Catalan.

Source Data

The original data comes from the Catalan subset of the Common Voice corpus.

Initial Data Collection and Normalization

Starting with version 13 of the Common Voice corpus we identified the speakers (273) who have recorded more than 1200 seconds of speech. A team of three annotators was tasked with annotating:

• if all the recordings correspond to the same person.
• the gender of the speaker
• the accent of the speaker
• the quality of the recording These annotations have been published in Zenodo.

They did a first round annotation, commented their different opinions and then, a second round. For this version of the corpus we used the annotations from the second round, and selected the IDs (263) where there was agreement between the annotators that correspond to a single person.

Who are the source language producers?

The original data comes from the Catalan sentences of the Common Voice corpus.

Annotations

From the original data in the Common Voice corpus the following fields have been added:

• annotated_gender(string): Annotated Gender by the experts team. It can be 'male' or 'female'

• annotated_gender_agreement (float): Agreement within the annotation team about the gender of the speaker.

• annotated_accent (string): Annotated Accent by the experts team. The accents considered are: Balearic, Central, Northern, Northwestern, Valencian.

• annotated_accent_agreement (float): Agreement within the annotation team about the accent of the speaker.

• mean quality (float): Mean annotated quality of the speakers' recording.

• stdev quality (float): Deviation in the quality of annotation between annotators.

• propagated_gender (string): Self-declared gender as indicated in certain recordings by the user. If there is a variation in the self-declared gender, "other" is assigned.

• propagated_accents (string): Self-declared accent as indicated in certain recordings by the user. If there is variation between different accents, the majority accent is propagated.

• propagated_accents_normalized (string): Propagated accent, normalized to the closed-options list used until version 7 (Balearic, Central, Northern, Northwestern, Valencian).

• assigned_accent (string): Accent assigned to the speaker. If the speaker has an expert annotation, it is prioritized. Otherwise, the propagated_accents_normalized option will be chosen. If expert annotation is not available and propagated_accents_normalized is not available either, the variant option is assigned. If none of these options are assigned, the field is left empty.

• assigned_gender (string): Gender assigned to the speaker. If the speaker has an expert annotation, it is prioritized. Otherwise, the propagated gender is assigned. If none of these options are assigned, the field is left empty.

Annotation process

273 speaker ID have been annotated. However, 10 of them have been discarded because they didn't correspond to a single individual.

The selected 263 speakers ID correspond to a total of 361,687 sentences in the validated split (20.18% of the total).

The full annotations made by the team of experts are available at Zenodo.

Who are the annotators?

The annotation was entrusted to the CLiC (Centre de Llenguatge i Computació) team from the University of Barcelona. They selected a group of three annotators (two men and one woman), who received a scholarship to do this work.

The annotation team was composed of:

• 2 male annotators, aged 18-25, L1 Catalan, students in the Catalan Philology degree.
• 1 female annotator, aged 18-25, L1 Catalan, student in the Modern Languages and Literatures degree, with a focus on Catalan.
• 1 female supervisor, aged 40-50, L1 Catalan, graduate in Physics and Linguistics, Ph.D. in Signal Theory and Communications.

To do the annotation they used a Google Drive spreadsheet

Personal and Sensitive Information

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

Considerations for Using the Data

Social Impact of Dataset

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

We hope that this corpus will be a useful contribution to the development of ASR tasks in Catalan

Discussion of Biases

Most of the voices of the common voice in Catalan correspond to men with a central accent between 40 and 60 years old. The aim of this dataset is to provide information that allows to minimize the biases that this could cause.

Regarding the content of the recorded sentences, we consider that the Common Voice validation system is efficient in removing those that could produce toxic content.

Other Known Limitations

[N/A]

Additional Information

Dataset Curators

Language Technologies Unit at the Barcelona Supercomputing Center ([email protected])

This work has been promoted and financed by the Generalitat de Catalunya through the Aina project.

Licensing Information

This dataset is licensed under a CC BY 4.0 license.

It can be used for any purpose, whether academic or commercial, under the terms of the license. Give appropriate credit, provide a link to the license, and indicate if changes were made.

Citation Information

DOI []

Contributions

The annotation was entrusted to the STeL team from the University of Barcelona.