Somali NLP research
Methodology and papers.
Every language needs AI infrastructure. We build it for Somali first, and to a standard the field can trust.
Documented and shareable
Datasets, model details, and evaluation methods are documented so others can verify and build on our releases.
Somali-led
Built with Somali speakers, who shape the technology and its direction.
Rigorous & reproducible
Clear benchmarks, documented methods, and results others can reproduce and extend.
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.
Speech Technology
ASR and TTS for Somali, turning radio, calls, and conversation into usable text, and text back into natural voice.
Foundation Models
Tokenizers and language models adapted to Somali morphology, dialects, and scripts, built to be efficient.
Research
Benchmarks, evaluation suites, and reproducible studies advancing low-resource NLP for African languages.
Research papers
Low-resource ASR for Somali: Corpus Design and Whisper Fine-tuning
Tood O., Yusuf S. · In preparation · 2026
Subword Tokenisation for Agglutinative Somali
Yusuf S., Tood O. · In preparation · 2025
New from the lab
Oct 27, 2026
Soplang 3.6: From Somali Programs to Native Applications
Soplang 3.6 adds files, networking, web servers, data clients, tasks, and project tools to the compiled Somali language. Official packages can now build native executables without installing Rust or Clang.
Oct 20, 2026
Somali tokenizer v2: 1.34 tokens per word on the 1.0.0 corpus
We scored our Somali ByteLevel BPE tokenizer on 63,542 held-out documents from Somali NLP Corpus 1.0.0. It needs 1.34 tokens per word against 1.81 for XLM-RoBERTa and 2.63 for BERT, it reconstructs every document exactly, and it is now on the Hugging Face Hub under Apache-2.0.
Oct 20, 2026
Releasing Somali NLP Corpus: 1.13 billion tokens
The first versioned release of our Somali corpus: 7,913,008 documents, 827 million words and 1.13 billion tokens from 14 public sources, with a license on every record, a gated Hugging Face dataset and an honest list of what is still wrong.
Oct 17, 2026
Releasing Soplang 3: Somali Syntax, Compiled Execution
Soplang 3 replaces the Python interpreter with a Rust compiler, Cranelift JIT and LLVM native builds. Explore its features, migration guide, and measured comparison with Soplang 2 and Python.
Sep 24, 2026
Somali NLP Corpus passes 1.1 billion tokens across 17 sources
17 upstream sources, 7.98M documents, 832M words and 1.136 billion subword tokens, measured over every document, plus what the four new sources changed and what the tokenizer still has to learn.
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 Somali AI with Somast
Use our published datasets and models, contribute to the research, or partner with us.
Researchers
Use our datasets and benchmarks, co-author studies, and push low-resource NLP forward.
Developers
Build on published models and tokenizers, file issues, and ship Somali-language features.
Students
Join fellowships and bootcamps to learn ML and contribute to real research.
Contributors
Help collect, transcribe, and verify data. Every contribution improves the models.
Need software built, or want to partner with us? Services · Partners