Spoken Somali, Afar, Arabic, and French become a filled-in form. A person checks it before it counts. Led by Radwan Jama.
Phoenix Rising Group turns spoken statements into draft forms for police desks, job centres, and weigh stations. An officer reviews every line before anything is saved. SASR Eval stays a research note.
Spoken Somali, Afar, Arabic, and French become a filled-in form. Emploi Assist, VoiceDoc, and Transport are the desks.
Record, draft, review, export. Nothing is saved until a person signs off.
Somali, Afar, Arabic, and French. Built for public service desks in Djibouti.
Public-data validation on 50 clips (~550 seconds) stays up as a long-game Somali ASR yardstick. The ASR lab track is frozen. Scores are not a homepage product claim. Full SASR Eval results.
| Component | Specification | Status |
|---|---|---|
| Base model | Whisper small (244M parameters), fine-tuned | Deployed |
| Training | ~1 hour on a single T4 GPU, 500 training steps | Complete |
| Public-data evaluation | SASR Eval: 50 clips, ~550 seconds, ground-truth WER/CER | Complete (June 2026) |
| Broadcast-domain evaluation | In-domain broadcast recordings from regional archives | Frozen / unpublished |
| Stage | Public-data yardstick. Live work is the three desks. | Research (frozen track) |
We are building infrastructure to process decades of broadcast data from the Horn of Africa including news, interviews, and cultural programming in Somali and regional languages. Archive work stays with Horn Transcribe. The desks are for forms a person checks.
Designed for continuous ingestion of Horn of Africa media from Djibouti, Somaliland, and adjacent territories where language technology has limited reach.
Infrastructure to digitize, transcribe, and index historical broadcast recordings. Enables retrospective analysis alongside ongoing archival work.
Detailed data access terms, archive specifications, and collection methodology are shared with qualified research and institutional partners under NDA.
Commercial systems underperform on Somali (WER 0.59–0.87 on the public evaluation set). Fine-tuned Whisper-small at WER 0.368 is the current public-data result.
Phoenix Rising Group is led by Radwan Jama, a software engineer with experience in applied machine learning and large-scale systems (Palo Alto Networks, Google Fellowship, Fidelity). He is building Phoenix Rising Group independently, focused on speech technology for Somali and other Horn of Africa languages.
The project combines:
Radwan studied Computer Science at Colgate University and has previously worked on ML-powered static analysis, large-scale systems, and AI security tooling. He built and validated SASR Eval end-to-end: dataset construction, model training, and benchmark scoring.
Universities, archives, and language programs focused on Horn of Africa languages and low-resource speech technology.
Groups digitizing broadcast, oral history, and cultural collections that need reliable Somali and regional-language transcription.
Public-sector and regional institutions exploring speech-enabled documentation workflows in Djibouti and the wider Horn of Africa.
Emploi Assist, VoiceDoc, and Transport are the desks. SASR Eval remains a research note.