kukuzaai/TTS-dataset

VoTexUg-TTS dataset

VoTexUg-TTS is a large-scale multilingual text-to-speech (TTS) dataset designed to support the development of high-quality speech synthesis models for widely spoken Ugandan languages. The dataset emphasizes linguistic diversity, gender balance, and natural African accents, addressing the persistent underrepresentation of African languages in speech technologies.

Dataset Details

Dataset Description

The VoTexUg-TTS dataset contains approximately 68 hours of studio-quality speech audio paired with text transcriptions. Data was collected by 12 students from Soroti University using a custom-built recording and upload system, ensuring consistency in audio quality and metadata capture. Each audio sample is sampled at 16 kHz and annotated with speaker identity, gender, and language.

To promote gender diversity, each of the six languages in the dataset is represented by two speakers (one male and one female). All recordings were reviewed and either approved or rejected through a structured quality assurance process before inclusion.

• Curated by: Kisejjere Rashid, Magala Reuben, Lynn Tar Gutu, Abubakhari Sserwadda
• Funded by: Soroti University Research and Innovation Fund (SUNRIF)
  https://sun.ac.ug/in-the-press/research-funds-call-for-proposals/
• Shared by: Soroti University
• Language(s) (NLP): English, Lusoga, Luganda, Runyankole, Acholi, Atteso
• License: More Information Needed

Dataset Sources

• Repository: More Information Needed
• Paper: More Information Needed
• Demo: More Information Needed

Uses

Direct Use

The dataset is intended for:

• Training and fine-tuning text-to-speech (TTS) models
• Research on multilingual and low-resource speech synthesis
• Accent-aware and African English speech generation
• Speaker-conditioned and gender-aware TTS systems

Out-of-Scope Use

The dataset is not suitable for:

• Speaker identification or biometric surveillance
• Emotion recognition tasks (emotions are not explicitly annotated)
• Automatic speech recognition (ASR) benchmarking without adaptation
• Any application involving impersonation or deceptive voice cloning

Dataset Structure

Each data sample consists of:

• audio: Speech waveform (16 kHz)
• text: Corresponding transcription
• speaker_id: Unique speaker identifier
• gender: Speaker gender (male or female)
• language: One of the six supported languages

The dataset is split into train and test sets. Splits were created to maintain balance across languages and genders.

Dataset Creation

Curation Rationale

The dataset was created to address the scarcity of high-quality TTS resources for Ugandan and African languages, and to enable the development of realistic voice generation models with authentic African accents.

Source Data

Data Collection and Processing

• A custom-designed web-based recording system was developed for the project
• Students recorded and uploaded speech data directly through the system
• Each submission underwent manual review and validation
• Recordings not meeting quality standards were rejected
• Accepted data was normalized and structured into standardized splits

Who are the source data producers?

The source data was produced by 12 student contributors from Soroti University, selected to represent native or fluent speakers of the target languages. Basic demographic metadata (language and gender) was collected; no personally identifiable information is included.

Annotations

Annotation process

Annotations include text transcriptions and speaker metadata (language, gender, speaker ID). Annotation and validation were carried out during the review stage using the custom system. No emotion or prosodic labels are provided.

Who are the annotators?

The annotations and quality reviews were performed by the project leads and designated reviewers from the Soroti University research team.

Personal and Sensitive Information

The dataset does not contain sensitive personal information. Speaker identities are anonymized using speaker IDs, and no names, addresses, or contact details are included.

Bias, Risks, and Limitations

• Limited number of speakers per language (two per language)
• Accents may reflect regional or educational backgrounds of student speakers
• Not all Ugandan languages are represented
• Dataset size may limit generalization for very large TTS models

Recommendations

Users should be aware of potential linguistic and demographic biases and are encouraged to:

• Combine this dataset with other African speech resources where possible
• Avoid misuse for identity inference or voice impersonation
• Clearly state dataset limitations when publishing results

Glossary

• TTS: Text-to-Speech
• Low-resource languages: Languages with limited digital and annotated data

Dataset Card Authors

Kisejjere Rashid, Magala Reuben, Lynn Tar Gutu, Abubakhari Sserwadda