[](https://aclanthology.org/2025.emnlp-main.559/)
[](https://africa.dlnlp.ai/simba/)
[](https://huggingface.co/spaces/UBC-NLP/SimbaBench)
[](https://github.com/UBC-NLP/simba)
[](https://huggingface.co/collections/UBC-NLP/simba-speech-series)
[](https://huggingface.co/datasets/UBC-NLP/SimbaBench_dataset)
## *Bridging the Digital Divide for African AI*
**Voice of a Continent** is a comprehensive open-source ecosystem designed to bring African languages to the forefront of artificial intelligence. By providing a unified suite of benchmarking tools and state-of-the-art models, we ensure that the future of speech technology is inclusive, representative, and accessible to over a billion people.
## Best-in-Class Multilingual Models
Introduced in our EMNLP 2025 paper *[Voice of a Continent](https://aclanthology.org/2025.emnlp-main.559/)*, the **Simba Series** represents the current state-of-the-art for African speech AI.
- **Unified Suite:** Models optimized for African languages.
- **Superior Accuracy:** Outperforms generic multilingual models by leveraging SimbaBench's high-quality, domain-diverse datasets.
- **Multitask Capability:** Designed for high performance in ASR (Automatic Speech Recognition) and TTS (Text-to-Speech).
- **Inclusion-First:** Specifically built to mitigate the "digital divide" by empowering speakers of underrepresented languages.
The **Simba** family consists of state-of-the-art models fine-tuned using SimbaBench. These models achieve superior performance by leveraging dataset quality, domain diversity, and language family relationships.
### 🔍 Simba-SLID (Spoken Language Identification)
* **🎯 Task:** `Spoken Language Identification` — Intelligent input routing.
* **🌍 Language Coverage (49 African languages)**
> **Akuapim Twi** (`Akuapim-twi`), **Asante Twi** (`Asante-twi`), **Tunisian Arabic** (`aeb`), **Afrikaans** (`afr`), **Amharic** (`amh`), **Arabic** (`ara`), **Basaa** (`bas`), **Bemba** (`bem`), **Taita** (`dav`), **Dyula** (`dyu`), **English** (`eng`), **Nigerian Pidgin** (`eng-zul`), **Ewe** (`ewe`), **Fanti** (`fat`), **Fon** (`fon`), **Pulaar** (`fuc`), **Pular** (`fuf`), **Ga** (`gaa`), **Hausa** (`hau`), **Igbo** (`ibo`), **Kabyle** (`kab`), **Kinyarwanda** (`kin`), **Kalenjin** (`kln`), **Lingala** (`lin`), **Lozi** (`loz`), **Luganda** (`lug`), **Luo** (`luo`), **Western Maninkakan** (`mlq`), **South Ndebele** (`nbl`), **Northern Sotho** (`nso`), **Chichewa** (`nya`), **Southern Sotho** (`sot`), **Serer** (`srr`), **Swati** (`ssw`), **Susu** (`sus`), **Kiswahili** (`swa`), **Swahili** (`swh`), **Tigre** (`tig`), **Tigrinya** (`tir`), **Tonga** (`toi`), **Tswana** (`tsn`), **Tsonga** (`tso`), **Twi** (`twi`), **Venda** (`ven`), **Wolof** (`wol`), **Xhosa** (`xho`), **Yoruba** (`yor`), **Standard Moroccan Tamazight** (`zgh`), **Zulu** (`zul`)
| **SLID Model** | **Architecture** | **Hugging Face Card** | **Status** |
| :--- | :--- | :---: | :---: |
| **Simba-SLID-49** 🔍 | HuBERT | 🤗 [https://huggingface.co/UBC-NLP/Simba-SLIS-49](https://huggingface.co/UBC-NLPSimba-SLIS-49) | ✅ Released |
**🧩 Usage Example**
You can easily run inference using the Hugging Face `transformers` library.
```python
from transformers import (
HubertForSequenceClassification,
AutoFeatureExtractor,
AutoProcessor
)
import torch
model_id = "UBC-NLP/Simba-SLIS_49"
model = HubertForSequenceClassification.from_pretrained(model_id).to("cuda")
# HuBERT models can use either processor or feature extractor depending on the specific model
try:
processor = AutoProcessor.from_pretrained(model_id)
print("Loaded Simba-SLIS_49 model with AutoProcessor")
except:
processor = AutoFeatureExtractor.from_pretrained(model_id)
print("Loaded Simba-SLIS_49 model with AutoFeatureExtractor")
# Optimize model for inference
model.eval()
audio_arrays = [] ### add your audio array
sample_rate=16000
nputs = processor(audio_arrays, sampling_rate=sample_rate, return_tensors="pt", padding=True).to("cuda")
# Different models might have slightly different input formats
try:
logits = model(**inputs).logits
except Exception as e:
# Try alternative input format if the first attempt fails
if "input_values" in inputs:
logits = model(input_values=inputs.input_values).logits
else:
raise e
# Calculate softmax probabilities
probs = torch.nn.functional.softmax(logits, dim=-1)
# Get the maximum probability (confidence) for each prediction
confidence_values, pred_ids = torch.max(probs, dim=-1)
# Convert to Python lists
pred_ids = pred_ids.tolist()
confidence_values = confidence_values.cpu().tolist()
# Get labels from IDs
pred_labels = [model.config.id2label[i] for i in pred_ids]
print(pred_labels, confidence_values)
```
## Citation
If you use the Simba models or SimbaBench benchmark for your scientific publication, or if you find the resources in this website useful, please cite our paper.
```bibtex
@inproceedings{elmadany-etal-2025-voice,
title = "Voice of a Continent: Mapping {A}frica{'}s Speech Technology Frontier",
author = "Elmadany, AbdelRahim A. and
Kwon, Sang Yun and
Toyin, Hawau Olamide and
Alcoba Inciarte, Alcides and
Aldarmaki, Hanan and
Abdul-Mageed, Muhammad",
editor = "Christodoulopoulos, Christos and
Chakraborty, Tanmoy and
Rose, Carolyn and
Peng, Violet",
booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2025",
address = "Suzhou, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.emnlp-main.559/",
doi = "10.18653/v1/2025.emnlp-main.559",
pages = "11039--11061",
ISBN = "979-8-89176-332-6",
}
```