Language Intelligence for the Horn of Africa
Andebet was founded in Addis Ababa to solve a fundamental technological gap: global automatic speech recognition systems fail on Ethiopian speech. We train neural acoustic models directly on authentic conversational audio across Ethiopian languages and regional dialects.
The Linguistic Reality
Ethiopia is home to over 80 languages, with Amharic, Afaan Oromo, and Tigrinya spoken by tens of millions of people in daily conversation, government administration, academic research, and national broadcasting. Yet commercial speech engines developed by major Western technology providers treat these languages as peripheral low-resource targets.
When international models attempt to transcribe an interview conducted in Addis Ababa, Hawassa, or Mekelle, they stumble on basic phonetics. Rapid switching between Amharic and English confuses tokenizers. Ethiopic fidel orthography requires precise distinction between 33 core consonants and seven vowel orders, as well as homophone normalization across characters like ሀ, ሐ, and ኀ. Off-the-shelf models produce disjointed approximations, missing critical timestamps and context.
Andebet addresses this challenge from the ground up. We collect, clean, and annotate authentic spoken audio recorded on real hardware—from mobile phone voice notes to field interviews conducted in noisy rural markets.
Neural Conformer Architecture
Our core transcription engine employs hybrid Conformer neural networks that combine self-attention mechanisms with depthwise convolution. Self-attention captures sentence-wide contextual relationships, while convolution models local speech dynamics such as gemination and vowel duration.
In rigorous evaluations on conversational Amharic benchmarks, our Conformer acoustic models achieve an 8.4% Word Error Rate (WER). The pipeline automatically normalizes Ge'ez homophones, preserves authentic Ethiopic punctuation marks including the two-dot word separator (፡) and four-dot full stop (።), and isolates speaker turns with millisecond accuracy.
Engineering & Operational Principles
Data Sovereignty
Field research, investigative journalism, and legal proceedings require uncompromising confidentiality. We offer isolated private cloud VPCs and on-premise deployments within Ethiopia.
Acoustic Verifiability
Every transcribed phrase links directly to its source audio timestamp. Researchers can click any Fidel word to verify speech playback, eliminating hallucinations in qualitative research.
Regional Coverage
We expand beyond capital city dialects. Our training corpora represent Gojjam, Wollo, Gondar, and Oromia acoustic environments to prevent demographic bias in speech AI.
Collaborate with our research team
Are you an academic institution, NGO, or media organization working with Ethiopian audio?