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.
Introduced in our EMNLP 2025 paper Voice of a Continent, 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.
The New Standard for African Speech-to-Text
🎯 Task Automatic Speech Recognition — Powering high-accuracy transcription across the continent.
🌍 Language Coverage (43 African languages)
Amharic (
amh), Arabic (ara), Asante Twi (asanti), Bambara (bam), Baoulé (bau), Bemba (bem), Ewe (ewe), Fanti (fat), Fon (fon), French (fra), Ganda (lug), Hausa (hau), Igbo (ibo), Kabiye (kab), Kinyarwanda (kin), Kongo (kon), Lingala (lin), Luba-Katanga (lub), Luo (luo), Malagasy (mlg), Mossi (mos), Northern Sotho (nso), Nyanja (nya), Oromo (orm), Portuguese (por), Shona (sna), Somali (som), Southern Sotho (sot), Swahili (swa), Swati (ssw), Tigrinya (tir), Tsonga (tso), Tswana (tsn), Twi (twi), Umbundu (umb), Venda (ven), Wolof (wol), Xhosa (xho), Yoruba (yor), Zulu (zul), Tamazight (tzm), Sango (sag), Dinka (din).
🌐 Explore the Frontier
| ASR Models | Architecture | #Parameters | 🤗 Hugging Face Model Card | Status |
|---|---|---|---|---|
| 🔥Simba-S🔥 | SeamlessM4T-v2 | 2.3B | 🤗 https://huggingface.co/UBC-NLP/Simba-S | ✅ Released |
| 🔥Simba-W🔥 | Whisper | 1.5B | 🤗 https://huggingface.co/UBC-NLP/Simba-W | ✅ Released |
| 🔥Simba-X🔥 | Wav2Vec2 | 1B | 🤗 https://huggingface.co/UBC-NLP/Simba-X | ✅ Released |
| 🔥Simba-M🔥 | MMS | 1B | 🤗 https://huggingface.co/UBC-NLP/Simba-M | ✅ Released |
| 🔥Simba-H🔥 | HuBERT | 94M | 🤗 https://huggingface.co/UBC-NLP/Simba-H | ✅ Released |
- Simba-S emerged as the best-performing ASR model overall.
🧩 Usage Example
You can easily run inference using the Hugging Face transformers library.
from transformers import pipeline
# Load Simba-S for ASR
asr_pipeline = pipeline(
"automatic-speech-recognition",
model="UBC-NLP/Simba-S" #Simba mdoels `UBC-NLP/Simba-S`, `UBC-NLP/Simba-W`, `UBC-NLP/Simba-X`, `UBC-NLP/Simba-H`, `UBC-NLP/Simba-M`
)
##### Load the multilingual African adapter (Only for `UBC-NLP/Simba-M`)
asr_pipeline.model.load_adapter("multilingual_african") # Only for `UBC-NLP/Simba-M`
###########################
# Transcribe audio from file
result = asr_pipeline("https://africa.dlnlp.ai/simba/audio/afr_Lwazi_afr_test_idx3889.wav")
print(result["text"])
# Transcribe audio from audio array
result = asr_pipeline({
"array": audio_array,
"sampling_rate": 16_000
})
print(result["text"])Using the same audio file with different Simba models:
# Simba-S
{'text': 'watter verontwaardiging sou daar, in ons binneste gewees het.'}# Simba-W
{'text': 'watter veronwaardigingsel daar, in ons binneste gewees het.'}# Simba-X
{'text': 'fator fr on ar taamsodr is'}# Simba-M
{'text': 'watter veronwaardiging sodaar in ons binniste gewees het'}# Simba-H
{'text': 'watter vironwaardiging so daar in ons binneste geweeshet'}Get started with Simba models in minutes using our interactive Colab notebook:
- 🎯 Task:
Text-to-Speech— Natural Voice Synthesis. 🌍 Language Coverage (7 African languages)
Afrikaans (
afr), Asante Twi (asanti), Akuapem Twi (akuapem), Lingala (lin), Southern Sotho (sot), Tswana (tsn), Xhosa (xho)
| TTS Model | Architecture | Hugging Face Card | Status |
|---|---|---|---|
| Simba-TTS-afr 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-afr | ✅ Released |
| Simba-TTS-twi-asanti 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-twi-asanti | ✅ Released |
| Simba-TTS-twi-akuapem 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-twi-akuapem | ✅ Released |
| Simba-TTS-lin 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-lin | ✅ Released |
| Simba-TTS-sot 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-sot | ✅ Released |
| Simba-TTS-tsn 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-tsn | ✅ Released |
| Simba-TTS-xho 🔊 | MMS-TTS | 🤗 https://huggingface.co/UBC-NLP/Simba-TTS-xho | ✅ Released |
🧩 Usage Example
You can easily run inference using the Hugging Face transformers library.
from transformers import VitsModel, AutoTokenizer
import torch
model_name="Simba-TTS-afr" ## Simba-TTS-twi-asanti, Simba-TTS-twi-akuapem, Simba-TTS-lin, Simba-TTS-sot, Simba-TTS-tsn, Simba-TTS-xho
model = VitsModel.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
text = "Ons noem hierdie deeltjies sub-atomiese deeltjies" #example of Afrikaans (afr) language
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
output = model(**inputs).waveformThe resulting waveform can be saved as a .wav file:
scipy.io.wavfile.write("outputfile.wav", rate=model.config.sampling_rate, data=output.float().numpy())Or displayed in a Jupyter Notebook / Google Colab:
from IPython.display import Audio
Audio(output.numpy(), rate=model.config.sampling_rate)
- 🎯 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 | ✅ Released |
🧩 Usage Example
You can easily run inference using the Hugging Face transformers library.
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)To evaluate your model on SimbaBench across all supported tasks (ASR, TTS, and SLID), simply load the corresponding configuration for the task and language you wish to benchmark.
Each task is organized by configuration name (e.g., asr_test_afr, tts_test_wol, slid_61_test). Loading a configuration provides the standardized evaluation split for that specific benchmark.
Example:
from datasets import load_dataset
data = load_dataset("UBC-NLP/SimbaBench_dataset", "asr_test_afr")DatasetDict({
test: Dataset({
features: ['split', 'benchmark_id', 'audio', 'text', 'duration_s', 'lang_iso3', 'lang_name'],
num_rows: 1000
})
})
data['test'][0]{'split': 'test',
'benchmark_id': 'afr_Lwazi_afr_test_idx3889',
'audio': {'path': None,
'array': array([ 4.27246094e-04, 7.62939453e-04, 6.71386719e-04, ...,
-3.05175781e-04, -2.13623047e-04, -6.10351562e-05]),
'sampling_rate': 16000},
'text': 'watter, verontwaardiging sou daar, in ons binneste gewees het?',
'duration_s': 5.119999885559082,
'lang_iso3': 'afr',
'lang_name': 'Afrikaans'}
| Config Name | Language | ISO | # Samples | # Hours |
|---|---|---|---|---|
| asr_test_Akuapim-twi | Akuapim-twi | Akuapim-twi | 1,000 | 1.35 |
| asr_test_Asante-twi | Asante-twi | Asante-twi | 1,000 | 0.97 |
| asr_test_afr | Afrikaans | afr | 1,000 | 0.87 |
| asr_test_amh | Amharic | amh | 581 | 1.12 |
| asr_test_bas | Basaa | bas | 582 | 0.76 |
| asr_test_bem | Bemba | bem | 1,000 | 2.15 |
| asr_test_dav | Taita | dav | 878 | 1.17 |
| asr_test_dyu | Dyula | dyu | 59 | 0.10 |
| asr_test_fat | Fanti | fat | 1,000 | 1.38 |
| asr_test_fon | Fon | fon | 1,000 | 0.66 |
| asr_test_fuc | Pulaar | fuc | 100 | 0.10 |
| asr_test_fuf | Pular | fuf | 129 | 0.03 |
| asr_test_gaa | Ga | gaa | 1,000 | 1.52 |
| asr_test_hau | Hausa | hau | 681 | 0.89 |
| asr_test_ibo | Igbo | ibo | 5 | 0.01 |
| asr_test_kab | Kabyle | kab | 1,000 | 1.05 |
| asr_test_kin | Kinyarwanda | kin | 1,000 | 1.50 |
| asr_test_kln | Kalenjin | kln | 1,000 | 1.50 |
| asr_test_loz | Lozi | loz | 399 | 0.91 |
| asr_test_lug | Ganda | lug | 1,000 | 1.65 |
| asr_test_luo | Luo (Kenya and Tanzania) | luo | 1,000 | 1.31 |
| asr_test_mlq | Western Maninkakan | mlq | 182 | 0.04 |
| asr_test_nbl | South Ndebele | nbl | 1,000 | 1.12 |
| asr_test_nso | Northern Sotho | nso | 1,000 | 0.88 |
| asr_test_nya | Nyanja | nya | 428 | 1.31 |
| asr_test_sot | Southern Sotho | sot | 1,000 | 0.82 |
| asr_test_srr | Serer | srr | 899 | 2.84 |
| asr_test_ssw | Swati | ssw | 1,000 | 0.93 |
| asr_test_sus | Susu | sus | 210 | 0.05 |
| asr_test_swa | Swahili | swa | 1,000 | 1.23 |
| asr_test_tig | Tigre | tig | 185 | 0.33 |
| asr_test_tir | Tigrinya | tir | 7 | 0.01 |
| asr_test_toi | Tonga (Zambia) | toi | 463 | 1.47 |
| asr_test_tsn | Tswana | tsn | 1,000 | 0.82 |
| asr_test_tso | Tsonga | tso | 1,000 | 0.99 |
| asr_test_twi | Twi | twi | 12 | 0.02 |
| asr_test_ven | Venda | ven | 1,000 | 0.92 |
| asr_test_wol | Wolof | wol | 1,000 | 1.19 |
| asr_test_xho | Xhosa | xho | 1,000 | 0.92 |
| asr_test_yor | Yoruba | yor | 359 | 0.42 |
| asr_test_zgh | Standard Moroccan Tamazight | zgh | 197 | 0.22 |
| asr_test_zul | Zulu | zul | 1,000 | 1.10 |
| Config Name | Language | ISO | # Samples | # Hours |
|---|---|---|---|---|
| tts_test_ewe | Ewe | ewe | 66 | 0.29 |
| tts_test_kin | Kinyarwanda | kin | 1,053 | 1.30 |
| tts_test_Asante-twi | Asante-twi | Asante-twi | 64 | 0.18 |
| tts_test_yor | Yoruba | yor | 40 | 0.13 |
| tts_test_wol | Wolof | wol | 4,001 | 4.12 |
| tts_test_hau | Hausa | hau | 124 | 0.24 |
| tts_test_lin | Lingala | lin | 63 | 0.28 |
| tts_test_xho | Xhosa | xho | 242 | 0.31 |
| tts_test_tsn | Tswana | tsn | 238 | 0.36 |
| tts_test_afr | Afrikaans | afr | 293 | 0.34 |
| tts_test_sot | Southern Sotho | sot | 210 | 0.33 |
| tts_test_Akuapim-twi | Akuapim-twi | Akuapim-twi | 83 | 0.22 |
| Config Name | Language Scope | # Samples | # Hours |
|---|---|---|---|
| slid_61_test | 61 Languages | 21,817 | 34.36 |
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.
@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",
}