Dataset Card for voxforge_spanish

Dataset Summary

VoxForge was set up to collect transcribed speech for use with Free and Open Source Speech Recognition Engines (on Linux, Windows and Mac). They promise they will make available all submitted audio files under the GPL license, and then 'compile' them into acoustic models for use with Open Source speech recognition engines such as CMU Sphinx, ISIP, Julius and HTK. According to this, we downloaded the Spanish recordings of Voxforge in 2016 to create the VOXFORGE SPANISH CORPUS.

The VOXFORGE SPANISH CORPUS has a duration of 49 hours and it is constituted by read speech recorded by more than 2 thousand speakers. Most of the speakers contribute with 10 recordings of approximately 10 seconds of duration each. Data is divided by speaker, by gender (male/female) and also by country (Argentina/Chile/LatinAmerica/Mexico/Spain/Unknown).

Example Usage

The VOXFORGE SPANISH CORPUS contains only the train split:

from datasets import load_dataset
voxforge_spanish = load_dataset("ciempiess/voxforge_spanish")

It is also valid to do:

from datasets import load_dataset
voxforge_spanish = load_dataset("ciempiess/voxforge_spanish",split="train")

Supported Tasks

automatic-speech-recognition: The dataset can be used to test 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 language of the corpus is Spanish.

Dataset Structure

Data Instances

{
  'audio_id': 'VXSP_F_0019_ARG_0033', 
  'audio': {
    'path': '/home/carlos/.cache/HuggingFace/datasets/downloads/extracted/7486eaf05a10c7554cd5de3d32c720fa206d11ad5f76e7f277553b34b1fbb58b/argentina/female/F_0019/VXSP_F_0019_ARG_0033.flac', 
    'array': array([-0.01412964, -0.02548218, -0.00692749, ..., -0.03274536,
       -0.03857422, -0.03134155], dtype=float32), 'sampling_rate': 16000
  }, 
  'speaker_id': 'F_0019', 
  'country': 'argentina', 
  'gender': 'female', 
  'duration': 11.5, 
  'normalized_text': 'todo estaba lo mismo que una hora antes cuando el té humeaba en la taza de ojeda ahora vacía y blanqueaban sobre la mesa los pliegos'
}

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
• country (string) - country of birth of the speaker.
• gender (string) - gender of speaker (male or female)
• duration (float32) - duration of the audio file in seconds.
• normalized_text (string) - normalized audio segment transcription

Data Splits

The corpus counts just with the train split which has a total of 21692 speech files from 467 female speakers and 1713 male speakers with a total duration of 49 hours and 42 minutes.

Dataset Creation

Curation Rationale

The VOXFORGE SPANISH CORPUS (VSC) has the following characteristics:

• The VSC has an exact duration of 49 hours and 42 minutes. It has 21692 audio files. 17053 of those files come from male speakers and 4639 come from female speakers. All recordings come from read speech.

• Male speakers contribute with 39h16m56s and female speakers contribute with 10h25m53s.

• The VSC counts with 2180 different speakers: 1713 men and 467 women.

• Every audio file in the VSC has a duration of 10 seconds approximately. Almost every speaker contributes with 10 recordings.

• Data in VSC is classified by speaker. It means, all the recordings of one single speaker are stored in one single directory.

• Data is also classified according to the gender (male/female) of the speakers and according to the nationality of the speaker (Argentina/Chile/LatinAmerica/Mexico/Spain/Unknown).

• Audio files in the VSC are distributed in a 16khz@16bit mono format.

• Information of an specific speaker can be tracked using the "Speaker_Info.xls" file to locate the Voxforge user name of that speaker and then, that user can be located in the Voxforge website.

• Every audio file has an ID that is compatible with ASR engines such as Kaldi and CMU-Sphinx.

Source Data

Initial Data Collection and Normalization

The VOXFORGE SPANISH CORPUS is a speech corpus designed to train acoustic models for automatic speech recognition and it is made out of recordings taken from VoxForge.

Annotations

Annotation process

The annotation process is at follows:

• 1. A whole podcast is manually segmented keeping just the portions containing good quality speech.
• 2. A second pass os segmentation is performed; this time to separate speakers and put them in different folders.
• 3. The resulting speech files between 5 and 10 seconds are transcribed by students from different departments (computing, engineering, linguistics). Most of them are native speakers but not with a particular training as transcribers.

Who are the annotators?

The VOXFORGE SPANISH CORPUS was created under the umbrella of the social service program "Desarrollo de Tecnologías del Habla" of the "Facultad de Ingeniería" (FI) in the "Universidad Nacional Autónoma de México" (UNAM) between 2016 and 2017 by Carlos Daniel Hernández Mena, head of the program.

Personal and Sensitive Information

The dataset could contain names revealing the identity of some speakers; on the other side, the recordings come from publicly available podcasts, so, there is not a real intent of the participants to be anonymized. Anyway, you agree to not attempt to determine the identity of speakers in this dataset.

Considerations for Using the Data

Social Impact of Dataset

This dataset is valuable because it contains well pronounced speech with low noise.

Discussion of Biases

The dataset is not gender balanced. It is comprised of 467 female speakers and 1713 male speakers.

Other Known Limitations

VOXFORGE SPANISH CORPUS by Carlos Daniel Hernández Mena is licensed under a GPLv3 license and it utilizes material from Voxforge. This work was done 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.

Dataset Curators

The dataset was collected by students belonging to the social service program "Desarrollo de Tecnologías del Habla". It was curated by Carlos Daniel Hernández Mena in 2017.

Licensing Information

GPLv3

Citation Information

@misc{carlosmena2017voxforgespanish,
      title={VOXFORGE SPANISH CORPUS: Audio and Transcriptions taken from Voxforge.org},
      author={Hernandez Mena, Carlos Daniel},
      organization={CIEMPIESS-UNAM Project},
      year={2017},
      url={https://huggingface.co/datasets/ciempiess/voxforge_spanish},
}

Contributions

The author would like to thank to Alejandro V. Mena, Elena Vera and Angélica Gutiérrez for their support to the social service program: "Desarrollo de Tecnologías del Habla." He also thanks to the social service students for all the hard work.

Special thanks to the VOXFORGE Team for publishing all the recordings that constitute the VOXFORGE SPANISH CORPUS.

This dataset card was created as part of the objectives of the 16th edition of the Severo Ochoa Mobility Program (PN039300 - Severo Ochoa 2021 - E&T).