The dataset stored in this repository is a collection of audio samples collected from a series of volunteers. Each .wav file contains voice samples from native Guyanese Creole (Creolese) speakers. The chosen volunteers come from a wide, but not exhaustive range of the Guyanese demographic, featuring speakers from Georgetown, Mahaica, Berbice, etc.

This dataset is intended for use in Automatic Speech Recognition (ASR) training, as well as other tasks such as:

Speech-to-text transcription
Accent and dialect modeling
Linguistic analysis of low-resource Caribbean languages
Fine-tuning pre-trained ASR models on Creolese data

The dataset was developed to support the representation and processing of Creolese in computational systems, addressing the underrepresentation of Caribbean English-lexified creoles in speech technology research.

How to Use:

This dataset is split into three folders:

finetune_eligible
finetune_eligible_draft
raw_audio_files

The format of a sample, and the value that sample can add to any future ASR tasks will determine which folder it will be stored in.

Below is a guide on how to add to and use this dataset.

./raw_audio_files

As the name suggests, this folder is where you will store the original copies of any sample you intend to add to the dataset. Only add a sample to this folder if you are certain that the speakers in the sample fall within the intended demographic of volunteers: Native Guyanese speakers, speaking in their comfortable register and code.

Adding an audio file to this folder will officially include that file in the corpora, however the presence of that sample in this folder does not mean it can be used for any processing tasks. To begin the preliminary preprocessing, run the following code in your command line:

Note: Only files of type .wav, .mp3, and .ogg can be added to this folder.

python chunk.py <name_of_auio_file.ogg>

Running chunk.py will:

Convert the file to .wav in 16000Hz, the preferred audio format of most ASR models.
Split that file into chunks of maximum 30 seconds in length, the preferred length of most ASR models.
Store the converted files in ../finetune_eligible_draft. If your audio is 30 seconds or shorter, it will be stored under it's original name. If your audio is longer, the chunks shall be concatenated with a letter addition from "a" to "zzz" to indicate sequence. E.g: 51a.wav, 51b.wav, ... , 51ax.wav, 51ay.wav

./finetune_eligible_draft

This folder consists of samples in the commonly used audio format for ASR tools: 30 seconds or shorter, .wav format, sampled at 16000Hz. This folder does not contain all chunked files. Once transcripts are provided for a sample, that sample is to be moved to finetune_eligible

./finetune_eligible

This folder consists of preprocessed samples and their accompanying transcripts, stored in transcripts.json. The file transcripts.json is a manifest of every sample that has a verified transcription. The structure of the data objects within are best exemplified below:

{

    "audio": "5l.wav",

    "text": "Wo kainda fruuts yuu laik? Wo kainda fuuds yuu laik? Somtaims aiz tink about gaiyaniiz fuuds. Dii kuk op an dii metemjii an dii korii. Bai a taim ai chraiyin fo mek korii. Bai? Ai likin miiself fo do korii. Ai put tuu moch spais insaid dis korii, nbadii neva tel mii dat diz hou yuz duu. Wach, ain gaa hombl miself an wach dem yuuchuub vidiiyoos an chrai mek it. Seem ting ai duu wid a kuk op. Bai a taim ai mek a kuk op. Ai mek di kuk op soo gud, it had piiz, it had edoo liivs, it had evriiting eksept teest. Bai ai fuget to pot dii salt insaid dis kuk op. And ai sorv it op nais nais, presenteeshon wan honjred.",

    "language": "crs"

}

In which:

"audio" represents the path to that sample
"text" represents the full transcript
"language" represents the language code ("crs" for Creoles)

Conclusion

As of 27th May, 2025, this dataset contains:

70 unprocessed samples in raw_audio_files.
38 untranscribed samples in finetune_eligible_draft.
265 transcribed files in `finetune_eligible.

length.py

To learn the total duration of the audio files in each folder, run the following:

python length,py

As of 27th May, 2025, there are:

174.74 minutes in raw_audio_files
16.51 minutes in finetune_eligible_draft
120.35 minutes in finetune_eligible Please update this section of README.MD as more samples are added to the dataset

Downloads last month: 6

Total file size:

462 MB

You need to agree to share your contact information to access this dataset

﻿# Creolese Audio Dataset