Research

Open research for Somali AI

How we do Somali NLP research — open methodology, reproducible results, peer-ready papers, and the SomBench evaluation suite.

Research philosophy

Every language deserves AI infrastructure. We build it for Somali first — openly, and to a standard the field can trust.

Open by default

Datasets, model weights, and evaluation code are released openly so anyone can build, verify, and improve them.

Community-owned

The people who speak the language help build the technology — and own its direction.

Rigorous & reproducible

Clear benchmarks, documented methods, and results others can reproduce and extend.

What we build

Four layers of the Somali AI stack

From raw data to deployable models and the research that holds it all to a standard.

Datasets

Open, permissively licensed Somali text, speech, and document corpora built for training and evaluation.

TextSpeechDocuments

Speech Technology

ASR and TTS for Somali — turning radio, calls, and conversation into usable text, and text back into natural voice.

ASRTTSAlignment

Foundation Models

Tokenizers and language models adapted to Somali morphology, dialects, and scripts — efficient and open.

LLMsTokenizersEmbeddings

Research

Benchmarks, evaluation suites, and reproducible studies advancing low-resource NLP for African languages.

BenchmarksEvaluationPapers
Publications

Research papers

BenchmarkEvaluationNLP

SomBench: A Benchmark for Somali Natural Language Processing

Yusuf S., Tood O. · In preparation · 2026

Abstract

We present SomBench, the first comprehensive evaluation benchmark for Somali NLP, covering classification, NER, translation, summarisation, and ASR. We evaluate six models and establish baselines for future work.

ASRSpeechWhisper

Low-resource ASR for Somali: Corpus Design and Whisper Fine-tuning

Tood O., Yusuf S. · In preparation · 2026

Abstract

We describe the construction of a 320-hour Somali speech corpus and show that fine-tuning Whisper on this data reduces WER by 41% over the zero-shot baseline, establishing a new state of the art on Somali ASR.

TokenizationMorphology

Subword Tokenisation for Agglutinative Somali

Yusuf S., Tood O. · In preparation · 2025

Abstract

We study the impact of BPE, Unigram, and morphology-aware tokenization on downstream Somali NLP tasks and show that morphology-informed models reduce out-of-vocabulary rates by 33% and improve classification F1 by up to 6 points.

SomBench v0.1

Benchmark results

Internal preview benchmarks on held-out data. Model weights and eval scripts releasing publicly soon.

TaskTypeMetricScoreModelStatus
ASR (clean)SpeechWER ↓
11.4
whisper-som-small preview
ASR (noisy)SpeechWER ↓
18.7
whisper-som-small preview
NERNLPF1
78.4
goobo-base preview
POS TaggingNLPAccuracy
86.2
goobo-base preview
SentimentNLPF1
82.1
goobo-base preview
SummarizationNLPROUGE-Lsom-sum upcoming
Translation (som→en)NLPBLEU
24.3
som-en-mt preview
Technical capabilities

The domains we work across

A full-stack research practice spanning language, speech, and retrieval — built for low-resource settings.

NLP

Tagging, parsing, NER

ASR

Speech to text

TTS

Text to speech

LLMs

Language models

Tokenizers

Subword for Somali

Benchmarks

Task suites

Evaluation

Rigorous metrics

RAG

Grounded retrieval

Speech Processing

Alignment & diarization

Build with us

Build Somali AI in the open

Use our datasets and models, contribute to the research, or partner with the lab. Everything we can open, we do.