Advancing Speech AI & Computational Linguistics for Ethiopian Languages
Empirical benchmarks, novel acoustic Conformer architectures, and open evaluation standards for Amharic, Afaan Oromo, Tigrinya, and low-resource African languages.
45 free minutes on signup · Telebirr & CBE Birr supported · Instant export to DOCX, SRT, VTT & JSON
Multi-Dialect Acoustic Evaluation (Wollo, Gojjam, Addis Ababa)
Amharic Dialectal Stress Test • Comparative Field Evaluation • 02:45
የገበያ ሁኔታውን ስናይ የጥራጥሬ እና የጤፍ ዋጋ ካለፈው ሳምንት ጀምሮ መጠነኛ መረጋጋት አሳይቷል።
“Looking at market conditions, the price of legumes and teff has exhibited moderate stabilization since last week.”
በዚህ በኩል ደግሞ የመስኖ ስራው በስፋት እየተካሄደ በመሆኑ የአትክልት ምርት በብዛት ወደ ገበያ እየገባ ይገኛል።
“On this side, because irrigation activities are proceeding extensively, vegetable yields are entering the market in high volume.”
የሞዴል ትክክለኛነት (CER): 97.2% | የድምጽ ዳራ ጫጫታ: SNR 14dB | የቃላት ስህተት ምጣኔ: 8.4%
“Model Confidence: CER 97.2% | Ambient Noise: SNR 14dB | Benchmark WER: 8.4%”
Core Pillars of Our Applied Research
Pioneering foundational speech recognition and natural language processing built specifically for the phonological and structural realities of Ethiopian languages.
Syllable-Aware Conformer Architectures
Replacing western subword BPE tokenizers with linguistically grounded Ge'ez syllabic vocabularies that accurately model consonant-vowel (CV) syllable blocks without token explosion.
Empirical Multi-Condition Benchmarks
Rigorous, reproducible evaluations measuring Word Error Rate (WER) and Character Error Rate (CER) across clean studio speech, telephonic bandwidths (8kHz), and noisy field recordings.
Morphological Complexity & OOV Mitigation
Semitic non-concatenative root-and-pattern templatic morphology creates millions of possible surface forms. Our language models use morphological decomposition to minimize out-of-vocabulary degradation.
Bilingual Code-Switching & Acoustic RAG
Developing neural representations that track rapid intra-sentential transitions between Amharic and English, enabling grounded conversational search and semantic retrieval across recorded audio archives.
Our Research & Model Development Pipeline
From ethical dialect collection and phonetic labeling to production Conformer deployment.
Corpus Curation & Ethical Collection
Curating authentic speech recordings across regional dialectal zones (Wollo, Gojjam, Gondar, Shewa, Oromia, Tigray) with full provenance and native speaker verification.
Over 15,000 hours of multi-genre audio
Phonetic Alignment & Orthographic Normalization
Resolving historic orthographic variants in Ge'ez fidel (such as ሀ/ሐ/ኀ, ሠ/ሰ, ዐ/አ, ጸ/ፀ) into standardized phonetic representations for unambiguous acoustic supervision.
Linguistically verified normalization rules
Joint CTC/Attention Conformer Training
Training high-capacity acoustic neural networks using depthwise separable convolutions and multi-head self-attention with dynamic masking and spec-augment perturbations.
Robust against background chatter and 8kHz phone audio
Stress Testing & Public Benchmark Reports
Validating models against out-of-domain conversational datasets, measuring degradation under varying signal-to-noise ratios (SNR), and publishing findings for the global NLP community.
Open methodology & reproducible metrics
Research Domains & Practical Applications
How Andebet's speech research translates into practical tools for engineering and fieldwork.
Speech Recognition Evaluation
AI Engineers & Research ScientistsAccess standardized benchmark methodology to evaluate Ethiopian speech engines against real-world conversational audio.
Computational Linguistics
Addis Ababa University & International NLP LabsIn-depth studies on Ge'ez fidel tokenization, phoneme duration modeling, and Semitic morphological decomposition.
Public Health & Humanitarian Fieldwork
WHO, UNICEF & Regional Health BureausHigh-accuracy transcription of regional health worker reports and community focus groups in low-connectivity settings.
Cultural Preservation & Oral History
National Archives & Cultural InstitutesPreserving fading oral traditions, folklore, and regional dialect recordings into indexed, searchable digital archives.
Frequently Asked Questions About Our Research
Insights into our acoustic models, benchmark protocols, and linguistic methods.
Ethiopian languages like Amharic and Tigrinya feature non-concatenative root-and-pattern morphology, an extensive 340+ glyph abugida script (Ge'ez fidel), gemination that carries semantic meaning but lacks explicit orthographic markers, and severe historical scarcity of clean training corpora. Afaan Oromo, while written in Latin Qubee, presents complex agglutinative morphology and phonemic vowel length distinctions that trip standard ASR models.
Because Amharic is morphologically rich (a single verb can contain subject, object, tense, and negation markers), a single mismatched suffix causes an entire word to count as an error in standard WER. We therefore report both WER and Character Error Rate (CER) after applying standardized Ge'ez orthographic normalization (collapsing phonetic duplicate characters).
Our training corpora specifically balance recordings from northern (Gondar, Wollo), western (Gojjam), central (Addis Ababa, Shewa), and southern regions. Our Conformer models learn shared acoustic representations that generalize across dialectal phonetic shifts and intonation differences.
In urban Ethiopian conversations, speakers routinely interleave English technical, financial, and colloquial terms into Amharic syntax. Our acoustic models are trained on real code-switched dialogues and dynamically emit both Ge'ez fidel and Latin characters with contextual language-model arbitration.
Yes! We actively collaborate with universities, computational linguists, and graduate researchers studying low-resource languages, African NLP, and acoustic modeling. Contact [email protected] to discuss data access and research partnerships.
Explore In-Depth Research Papers & Benchmarks
Amharic Speech Recognition Benchmark
Full empirical benchmark: WER, CER, dialect evaluations, and noise resistance.
WhitepaperEthiopian Language AI & Low-Resource NLP
Computational linguistics paper on Ge'ez syllabic tokenization and morphological modeling.
ProductAmharic Speech to Text
Production API and transcription software powered by our research models.
PlatformCorpus Analysis & Audio RAG
Interactive semantic search and question answering across multi-hour audio collections.
Turn recorded Ethiopian conversations into searchable truth.
Join researchers, media networks, and global organizations extracting structured knowledge from Amharic, Afaan Oromo, and Tigrinya audio.
- 45 free minutes included
- No credit card required
- Instant API & Workspace access
- Enterprise data sovereignty