license: cc0-1.0
dataset_info:
features:
- name: filename
dtype: string
- name: gender
dtype: string
- name: age
dtype: string
- name: language
dtype: string
- name: text
dtype: string
- name: audio
dtype: audio
splits:
- name: train
num_bytes: 1047469197.184
num_examples: 5242
download_size: 879436814
dataset_size: 1047469197.184
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
CommonPhone-SE
Multilingual, age and gender balanced subset for speech enhancement benchmark.
Dataset Details
Dataset Description
Commonphone-SE is a benchmark dataset derived from Commonphone. It contains audio samples from 7 languages in the age range from 18 to 80. It aims to provide a speaker diverse dataset to benchmark speech enhancement algorithms in real world conditions.
- Curated by: LangTech Lab members from the speech team.
- Language(s) (NLP): CA, DE, EN, IT, FR, RU, ES
- License: cc0-1.0
Dataset Sources
CommonPhone-SE is a subset of CommonPhone.
- Repository: https://zenodo.org/records/5846137
- Paper : https://arxiv.org/abs/2201.05912
Languages
Catalan(CA), Deutsch(DE), English(EN), Italian(IT), French(FR), Russian(RU), Spanish(ES)
Uses
The goal of this dataset is to evaluate the generalization capabilities of speech enhancement models in a real world multilingual and diverse dataset.
Dataset Structure
The dataset consists of a single split, providing audios, transcriptions and demographic information
Dataset({
features: ['filename', 'gender', 'age', 'language', 'text', 'audio'],
num_rows: 5242
})
Each data point is structured as:
{'filename': 'common_voice_ca_31498257',
'gender': 'female',
'age': 'fifties',
'language': 'ca',
'text': 'lieutenant monroe va resultar ferit durant la batalla i va servir posteriorment al congrés',
'audio': {'path': 'Commonphone-SE/common_voice_ca_31498257.wav',
'array': array([0., 0., 0., ..., 0., 0., 0.]),
'sampling_rate': 16000}}
Dataset Creation
Curation Rationale
The sampling rationale was to select audios that remain difficult for state of the art enhancement models, both in terms of speech quality metrics and content preservation, hence, we selected the worst 40 examples w.r.t. to UTMOS, SCOREQ and WIL per each language, age band and gender. Finally, the duplicates were dropped to arrive at a final evaluation dataset of 8.24 hours.
Source Data
Crowdsourced audios recorded by volunters for CommonVoice that were selected in the CommonPhone dataset.
Data Collection and Processing
Who are the source data producers?
Common Phone is maintained and distributed by speech researchers at the Pattern Recognition Lab of Friedrich-Alexander-University Erlangen-Nuremberg (FAU)
Personal and Sensitive Information
Like for Common Voice, you must not make any attempt to identify speakers that contributed to CommonPhone-SE.
Bias, Risks, and Limitations
The dataset was built trying to mitigate the bias on gender and age variables, however, it can still be biased towards the degradations found in the commonvoice corpus. Althoug this dataset has a lot of diversity the style is only reading speech.
Citation
If you use the dataset please cite
BibTeX:
@inproceedings{giraldo25_interspeech,
title = {{Evaluating Speech Enhancement Performance Across Demographics and Language}},
author = {{Jose Giraldo and Alex Peiró-Lilja and Carme Armentano-Oller and Rodolfo Zevallos and Cristina España-Bonet}},
year = {{2025}},
booktitle = {{Interspeech 2025}},
pages = {{1353--1357}},
doi = {{10.21437/Interspeech.2025-1760}},
issn = {{2958-1796}},
}
Dataset Card Authors
Funding
This work has been promoted and financed by the Generalitat de Catalunya through the Aina project and also by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación,Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215337