Building for Ethiopia: language tools, AI models, and speech datasets

Products serving Ethiopian users run into a challenge early: Amharic uses the Ethiopic script (Fidel), a syllabic alphabet where a single character encodes a consonant-vowel pair. Standard NLP pipelines built for Latin-script languages break at the tokenization step. Character normalization, morphological analysis, and sentence segmentation all require tooling built specifically for Ethiopic text.

The Wycord registry has 16 verified resources for Ethiopia. Payments, SDKs, and the Chapa MCP server are covered in the Ethiopia payments guide. This guide covers the rest: three NLP tools, one production translation API, one open language model family, and six speech and ASR datasets. Together they form the non-payment stack for building in Ethiopia.

Amharic text processing: three complementary tools

Three open-source Python tools in the registry each handle a different layer of Amharic text. Using them together covers the full preprocessing pipeline.

amseg works at the document level. It segments Ethiopic text into sentences and tokens, normalizes character variants across Unicode codepoints (Ethiopic has multiple representations for some characters that look identical), and transliterates bidirectionally between Fidel and Latin scripts. The Latin transliteration is useful when you need to pass Amharic text through a downstream model that does not support Ethiopic input natively. Developed at the University of Hamburg. GitHub: uhh-lt/amharicprocessor.

etnltk sits above amseg in the stack. It handles text normalization, short-form expansion, and word and sentence tokenization for Amharic and related Ethiopian languages. The API follows spaCy and NLTK conventions, so developers already using those libraries will recognize the patterns. Available on PyPI as etnltk. GitHub: robeleq/etnltk.

HornMorpho goes deeper than tokenization. It performs morphological analysis and generation for Amharic, Oromo, and Tigrinya: it breaks words into their constituent morphemes and can synthesize word forms from roots and grammatical structures. This matters for search indexing (morpheme-level indexing catches inflected variants), NER pipelines that need to normalize proper nouns across case and gender inflections, and any downstream task where word structure carries meaning. Originated at Indiana University's L3 Project. GitHub: fgaim/HornMorpho.

A typical preprocessing pipeline for Amharic: amseg to normalize and segment the document, etnltk to tokenize and expand abbreviations, HornMorpho for tasks requiring morphological analysis.

Machine translation: Lesan AI

Lesan AI is the only production translation API in the Ethiopia registry. It covers Amharic, Tigrinya, Oromo, and Somali, with a focus on language pairs where major providers perform poorly. Founded in Ethiopia in 2019.

Pricing is EUR 20 per million characters, consumption-based with no subscription. The API endpoint is documented at lesan.ai/services.html.

For products that need to translate UI strings, support content, or user messages across Ethiopian language pairs, Lesan AI is the verified starting point. It outperforms generic translation services on Ethiopian languages because the training data and focus are specific to those languages, not incidental coverage from a multilingual model trained primarily on European text.

Open language model: EthioLLM

EthioLLM is a multilingual language model family developed by the EthioNLP collective, built on XLM-RoBERTa-Large. It covers six languages: Amharic, Ge'ez, Afaan Oromoo, Somali, Tigrinya, and English. The large variant is available on HuggingFace as EthioNLP/EthioLLM-l-70K under an MIT license.

EthioLLM is a masked language model, which makes it a strong base for fine-tuning on classification, named entity recognition, and sentiment analysis tasks in Ethiopian languages. Standard HuggingFace loading works:

from transformers import pipeline
pipe = pipeline("fill-mask", model="EthioNLP/EthioLLM-l-70K")

The 2024 release paper (Tonja et al., arXiv:2403.13737) publishes benchmark accuracy on classification and NER tasks across all five Ethiopian languages, giving you a baseline to measure against when fine-tuning.

This is the only open Ethiopian language model in the Wycord registry with a clean license and documented benchmarks. If you are fine-tuning rather than using a hosted API, start here.

Speech datasets: six Amharic resources

Six speech and ASR datasets for Amharic are verified in the registry. All are hosted on HuggingFace. None currently carry a formal open-data license in the registry records; confirm usage rights with the dataset authors before building a production system on them.

Amharic Speech (snapwre/amharic-speech): a crowdsourced Amharic speech dataset for training speech processing models. Community-submitted; carries a "needs review" flag in the Wycord registry.

Amharic ASR Benchmark (snapwre/amharic-asr-benchmark): a held-out benchmark set for evaluating ASR systems on Amharic. Use this to score a trained model rather than as training data. Also community-submitted; needs review flag applies.

Leyu Amharic dialect series: four separate datasets covering distinct Amharic-speaking regions: Addis Ababa, Gojjam, Gonder, and Wello. The dialect split matters in practice. Amharic pronunciation and vocabulary vary across regions, and a model trained only on Addis Ababa speech will underperform on Gojjam or Gonder speakers. All four are in the leyu-amharic HuggingFace organization.

For teams building voice-first products or ASR features for Ethiopia, using multiple dialect datasets during training, rather than a single regional corpus, is the more robust approach.

What to use when

Task	Verified option
Amharic Unicode normalization, sentence segmentation	amseg
Latin to Fidel (or Fidel to Latin) transliteration	amseg
Amharic text normalization, tokenization, abbreviation expansion	etnltk
Morphological analysis of Amharic, Oromo, or Tigrinya	HornMorpho
Machine translation across Amharic, Tigrinya, Oromo, Somali	Lesan AI (EUR 20/1M chars)
Fine-tuning for Amharic or Ethiopian language NER, classification, sentiment	EthioLLM (MIT, HuggingFace)
Training an Amharic ASR model	Amharic Speech (community; verify license)
Evaluating an Amharic ASR model	Amharic ASR Benchmark (community; verify license)
Dialect-specific ASR or TTS training data	Leyu dialect series (4 regions)
Payment acceptance, mobile wallet integration	Chapa, ArifPay: see the payments guide

The full list of all 16 verified Ethiopia resources, with live status indicators, is at wycord.com/countries.