Dataset Card for dimex100_light

Dataset Summary

The DIMEx100 LIGHT Corpus (DL) is a reduced version of the DIMEx100 Corpus (D100). DL was created in 2016 by Carlos Daniel Hernández Mena, with the aim of facilitating the use of the DIMEx100 Corpus in various automatic speech recognition systems.

The most important differences between DIMEx100 LIGHT and the original are:

• The DL only contains audio files and transcriptions, unlike the D100 which contains pronunciation dictionaries, timestamps, and phonetic labeling with different levels of granularity.

• The DL's directory organization efficiently presents information about the speaker's gender and the type of utterance. This same information is also present in the file names of each audio in the corpus. All of this allows for greater convenience when creating corpus subsets for both training and testing through computational scripts.

The most important features that DL and D100 have in common are:

• The folder names with audio in D100 start with "s", from "s001" to "s100". These same folder names are retained by the DL, although the audios are named in a way that is not identical but equivalent. Within all the audio folders, the DL also maintains the distinction between "common" and "individual" utterances, just like the D100.

• The audios in both corpora have a format of 16 bits, mono, with a sampling rate of 16 kHz. In the case of the DL, the audios were normalized using SoX.

Example Usage

The DIMEx100 LIGHT contains only the train split:

from datasets import load_dataset
dimex100_light = load_dataset("carlosdanielhernandezmena/dimex100_light")

It is also valid to do:

from datasets import load_dataset
dimex100_light = load_dataset("carlosdanielhernandezmena/dimex100_lightt",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 with the accent of Central Mexico.

Dataset Structure

Data Instances

{
  'audio_id': 'DXCL_F_040_I_007', 
  'audio': {
    'path': '/home/carlos/.cache/HuggingFace/datasets/downloads/extracted/d6b730265a2ff71f70a18e25c1b7b74d947cb8626af39b58806ad160638707ab/train/female/s040/individuales/DXCL_F_040_I_007.flac', 
    'array': array([ 5.1269531e-03,  9.1247559e-03,  7.5683594e-03, ...,
        9.1552734e-05, -2.7465820e-04, -3.6621094e-04], dtype=float32), 
    'sampling_rate': 16000
  },
  'speaker_id': 'F_040', 
  'utterance_type': 'individual', 
  'gender': 'female', 
  'duration': 4.61899995803833, 
  'normalized_text': 'informe sobre el uso de normas para la transmisión de señales de televisión'
}

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
• utterance_type (string) - indicates whether the utterance is part of the set that all the speakers read (common) or is part of the set that is unique to each speaker (individual).
• 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 5999 speech files from 50 male speakers and 50 female speakers with a total duration of 6 hours and 9 minutes.

Dataset Creation

Curation Rationale

The DIMEx100 LIGHT (DL) Corpus has the following characteristics:

• The DL has a total of 5999 audio files of 50 male speakers and 50 female speakers. It has a total duration of 6 hours and 9 minutes.

• The total number of audio files that come from male speakers is 3000 with a total duration of 2 hours and 57 minutes. The total number of audio files that come from female speakers is 2999 with a total duration of 3 hours and 12 minutes. So, DL is perfectly balanced in gender.

• Every audio file in the DL has a duration between 2 and 10 seconds approximately.

• Data in DL is classified by gender and also by speaker, so one can easily select audios from a particular set of speakers to do experiments.

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

Source Data

Initial Data Collection and Normalization

The corpus was created from sentences selected from the Web, all of them ordered from least to most complex, ensuring they were always phonetically balanced.

Annotations

Annotation process

The distribution of the number of sentences was selected as follows: each of the 100 speakers read a total of 50 sentences, resulting in a total of 5,000 recorded sentences. Each speaker additionally recorded 10 common sentences, totaling 1,000 additional sentences; therefore, the DIMEx100 corpus is composed of 6,000 sentences.

Who are the annotators?

The DIMEx100 Corpus was created in the "Departamento de Ciencias de la Computación" (DCC) at the "Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas" (IIMAS) of the "Universidad Nacional Autónoma de México" (UNAM) as part of the DIME-II project.

Personal and Sensitive Information

The dataset does not contain names revealing the identity of any of the speakers. However, 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 pristine read speech from speakers from Central Mexico.

Discussion of Biases

The characteristics of the speakers selected for the recordings of these corpora were primarily age, education level, and place of origin. According to these characteristics, the speakers who participated in the creation of the DIMEx100 Adults corpus were from central Mexico, specifically Mexico City; most were students, researchers, and professors from UNAM. The age range of the speakers was 16 to 32 years, with an average age of 24. Regarding gender, 49% of the speakers were men and 51% were women.

Other Known Limitations

"DIMEx100 LIGHT CORPUS" by Carlos Daniel Hernández Mena is licensed under a Attribution-NonCommercial-NoDerivatives 4.0 International (CC-BY-NC-ND 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.

Dataset Curators

The dataset was curated by Carlos Daniel Hernández Mena in 2016 from the original DIMEx100 Corpus created in 2004 at IIMAS-UNAM (Mexico City).

Licensing Information

CC-BY-NC-ND-4.0

Citation Information

@misc{carlosmena2016dimex100light,
      title={DIMEx100 LIGHT CORPUS: Reduced version of the DIMEx100 Corpus created in 2004 at IIMAS-UNAM.}, 
      author={Hernández Mena, Carlos Daniel},
      organization={CIEMPIESS-UNAM Project},
      year={2016},
      url={https://huggingface.co/datasets/carlosdanielhernandezmena/dimex100_light},
}

Contributions

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).