Professor/fongbe-speech-zenodo

Fongbe Speech Dataset (Complete & Tone-Preserved)

Dataset Summary

This dataset is a unified, high-quality collection of Fongbe speech data, specifically curated to preserve the linguistic integrity of this tonal language. It acts as a complete, unsegmented, and tone-accurate assembly of the Fongbe Continuous Speech Recognition corpora, merging:

1. The foundational ALFFA Project data (Train/Test splits, 2016).
2. The expanded Zenodo release (Validation split, 2022).

Why Use This Version?

Several derivative Fongbe datasets currently on the Hub have been heavily pre-processed. Those versions often chunk audio into unnatural segments, remove punctuation, or strip critical tone diacritics.

This repository ensures linguistic accuracy by preserving:

• Tones: Retains all diacritics (e.g., ɖ, ɛ, ɔ, è, é, ì, ̌, ̂, ĕ, ŏ) essential for Fongbe semantics.
• Audio Integrity: Provides full-length original utterances rather than aggressively chopped segments, allowing for better context modeling.
• Harmonized Schema: Standardizes metadata across diverse source origins for immediate use in transformers.

Dataset Statistics

Split	Utterances	Total Duration (approx)	Source
Train	8,234	~5.73 hrs	ALFFA GitHub
Validation	3,179	~5.11 hrs	Zenodo (6604637)
Test	2,168	~1.45 hrs	ALFFA GitHub
Total	13,581	~12.30 hrs

Technical Metadata

• Sampling Rate: 16,000 Hz
• Audio Format: WAV (PCM 16-bit)
• Language: Fongbe (Fon)
• Tonal Representation: Decomposed (NFD) diacritics preserved.

Dataset Structure

All splits share a harmonized schema for seamless use with the Hugging Face datasets and transformers libraries.

Features

• audio: A dictionary containing the audio array, sampling rate (16kHz), and local path.
• text: The ground-truth transcription with full tonal markers and punctuation.
• speaker ID: The unique identifier for the speaker (e.g., "denise", "mario", "5").
• audio filename: The original filename for provenance tracking.

Quick Start

You can load this dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

# Load the full DatasetDict
dataset = load_dataset("Professor/fongbe-speech-zenodo")

# Access individual splits
train_data = dataset["train"]
validation_data = dataset["validation"]
test_data = dataset["test"]

# View a sample entry
print(train_data[0])

Sources & Credits

This dataset is the result of extensive fieldwork and research by Fréjus A. A. Laleye and the ALFFA (African Languages in the Field) project team.

Citation Information

If you use this dataset, please cite the following works:

For the Training/Test Data (ALFFA):

@inproceedings{laleye2016FongbeASR,       
    title={First Automatic Fongbe Continuous Speech Recognition System: Development of Acoustic Models and Language Models},     
    author={A. A Laleye, Fréjus and Besacier, Laurent and Ezin, Eugène C. and Motamed, Cina},     
    year={2016},     
    organization={Federated Conference on Computer Science and Information Systems}
}

For the Validation Data (Zenodo):

@dataset{laleye_frejus_2022_6604637,
  author       = {Laleye, Fréjus A. A.},
  title        = {Fongbe Speech Dataset},
  month        = jun,
  year         = 2022,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.6604637},
  url          = {https://doi.org/10.5281/zenodo.6604637}
}

Contributions

Ported to Hugging Face by Victor Olufemi (Professor). This version ensures that low-resource language modeling for Fongbe remains linguistically accurate by preventing the loss of tonal information during preprocessing.