Sukuma Voices 🎙️

The first publicly available speech corpus for Sukuma (Kisukuma), a Bantu language spoken by approximately 10 million people in northern Tanzania.

---

📊 Dataset Overview

Metric	Value
Total Samples	6,871
Total Duration	19.56 hours
Average Duration	10.25 ± 4.15 seconds
Duration Range	1.40 - 30.36 seconds
Total Words	140,325
Unique Vocabulary	21,366
Average Words/Sample	20.4
Speaking Rate	121.6 WPM

🎯 Supported Tasks

• Automatic Speech Recognition (ASR) — Converting Sukuma speech to text
• Text-to-Speech (TTS) — Synthesizing natural-sounding Sukuma speech
• Cross-lingual Speech Processing — Research between Swahili and Sukuma

🚀 Quick Start

Installation

pip install datasets transformers librosa

Load the Dataset

from datasets import load_dataset

# Load the dataset from HuggingFace
dataset = load_dataset("sartifyllc/Sukuma-Voices", split='train')

# View a sample
print(dataset[0])

ASR Inference Example

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

# Load model and processor
model = WhisperForConditionalGeneration.from_pretrained("sartifyllc/sukuma-voices-asr")
processor = WhisperProcessor.from_pretrained("sartifyllc/sukuma-voices-asr")

# Load and preprocess audio
audio_array = ...  # Your audio as numpy array at 16kHz

input_features = processor(
    audio_array, 
    sampling_rate=16000, 
    return_tensors="pt"
).input_features

# Generate transcription
with torch.no_grad():
    predicted_ids = model.generate(input_features)

# Decode
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

📁 Repository Structure

sukuma-voices/
├── README.md
├── LICENSE
├── scripts/
    ├── train_asr.py
    └── train_tts.py

📈 Baseline Results

ASR Performance (Whisper Large V3)

Metric	Original Speech	Synthetic Speech
Final WER	25.19%	32.60%
Min WER	22.01%	29.97%
WER Reduction	82.94%	78.93%

TTS Performance (Orpheus 3B v0.1)

Metric	Score
Mean Opinion Score (MOS)	3.9 ± 0.15
Human Recording MOS	4.6 ± 0.1

📝 Example Data

Language	Text
Sukuma	Umunhu ngwunuyo agabhalelaga chiza abhanhu bhakwe, kunguyo ya kikalile kakwe akagubhatogwa na gubhambilija abho bhali mumakoye.
English	This person raises his people well, because of his good behavior, of loving people and helping his colleagues who are in trouble, in their lives.

🔬 Reproducing Results

Training ASR

export WANDB_API_KEY="your_key"
export HF_TOKEN="your_token"
python train_asr.py

Training TTS

export WANDB_API_KEY="your_key"
export HF_TOKEN="your_token"
python scripts/train_tts.py 

⚠️ Known Limitations

• Domain Specificity: Data sourced from biblical texts may not fully represent everyday conversational Sukuma
• Diacritic Variations: Sukuma has two written forms; this dataset uses the non-diacritic version
• Speaker Diversity: Limited speaker diversity from a single recording source

📜 Citation

If you use this dataset in your research, please cite:

@inproceedings{mgonzo-etal-2026-learning,
    title = "Learning from Scarcity: Building and Benchmarking Speech Technology for {S}ukuma.",
    author = "Mgonzo, Macton  and
      Oketch, Kezia  and
      Etori, Naome A  and
      Mang'eni, Winnie  and
      Nyaki, Elizabeth Fabian  and
      Mollel, Michael Samwel",
    editor = "Hettiarachchi, Hansi  and
      Ranasinghe, Tharindu  and
      Plum, Alistair  and
      Rayson, Paul  and
      Mitkov, Ruslan  and
      Gaber, Mohamed  and
      Premasiri, Damith  and
      Tan, Fiona Anting  and
      Uyangodage, Lasitha",
    booktitle = "Proceedings of the Second Workshop on Language Models for Low-Resource Languages ({L}o{R}es{LM} 2026)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.loreslm-1.25/",
    doi = "10.18653/v1/2026.loreslm-1.25",
    pages = "288--296",
    ISBN = "979-8-89176-377-7",
    abstract = "Automatic Speech Recognition (ASR) systems are gaining increasing attention in both academia and industry. Despite having remarkable performance in high-resource languages, their efficacy is less pronounced in low-resource settings. We present the first ASR system for Sukuma, one of the most severely under-resourced Tanzanian languages, and provide an open-source Sukuma speech corpus comprising 7.47 hours of carefully transcribed audio. The data, sourced primarily from Bible readings, was rigorously annotated to ensure phonetic and orthographic consistency, making it the most linguistically reliable resource currently available for the Sukuma language. To establish baselines, we train lightweight ASR and Text-to-Speech (TTS) models that demonstrate the feasibility of building end-to-end speech systems for this underrepresented language. This work addresses the challenges of developing language and communication tools for speakers of less-represented languages, particularly the scarcity of representative datasets and benchmarks, and highlights future research directions for linguistically challenging languages, such as Sukuma. We make our data and code publicly available to facilitate reproducibility and further research."
}

👥 Authors

Name	Affiliation	Contact
Macton Mgonzo	Brown University	macton_mgonzo@brown.edu
Kezia Oketch	University of Notre Dame
Naome Etori	University of Minnesota - Twin Cities
Winnie Mang'eni	Pawa AI
Elizabeth Nyaki	Pawa AI, Sartify Company Limited
Michael S. Mollel	Pawa AI, Sartify Company Limited	michael.mollel@sartify.com

🙏 Acknowledgments

We would like to express our gratitude to Sartify Company Limited and Pawa AI for their instrumental role in initiating this project and for providing the data access necessary to develop and evaluate our models. We also extend our sincere thanks to all the volunteers who generously dedicated their time to the evaluation process.

📄 License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

🤝 Contributing

We welcome contributions to expand and improve this dataset! Areas of interest include:

• Additional Sukuma speech data beyond religious content
• Conversational and everyday language recordings
• Multi-speaker recordings
• Diacritic-annotated transcriptions

Please open an issue or submit a pull request to contribute.

📧 Contact

For questions, collaborations, or feedback, please:

• Open an issue on this repository
• Contact: macton_mgonzo@brown.edu, info@sartify.com

---

This dataset represents an important step toward inclusive speech technology for African languages.

Sartify • Pawa AI • HuggingFace