Language Technology · Research Stage Somali ASR

Preserving Horn of Africa Languages Through Speech AI

Phoenix Rising Group builds Somali speech recognition technology validated against commercial systems. 37.8% lower word error rate than off-the-shelf ASR.

We process decades of Horn of Africa broadcast, oral history, and cultural archives so that Somali, Afar, and other regional languages are represented in the AI systems being built today.

View SASR Eval Results Learn More

SASR Eval: Mean Word Error Rate

Fine-tuned Whisper small (ours) 0.368

ElevenLabs Scribe v1 0.592

faster-whisper large-v3 0.871

Ours ElevenLabs Whisper

Evidence

Validated performance, not marketing claims.

SASR Eval is the first ground-truth evaluation of commercial and fine-tuned ASR systems on Somali speech. Fifty stitched clips, ~550 seconds total, scored against verified references.

0.368

Mean WER
Fine-tuned Whisper small

37.8%

Relative WER reduction
vs. ElevenLabs Scribe v1

57.8%

Relative WER reduction
vs. Whisper large-v3

Evaluated clips
~10 seconds each

What we do

Language preservation through speech AI.

Phoenix Rising Group is building a platform to process and preserve broadcast, oral history, and cultural archives from Djibouti, Somaliland, and the broader Horn of Africa.

Archive digitization

Automated ingestion and transcription of regional broadcast media, oral history, and cultural programming for long-term preservation.

Domain-specific ASR

Our fine-tuned pipeline outperforms commercial systems on Somali speech. Trained in ~1 hour on a single T4 GPU with 500 steps.

Regional access

Father-son team with deep cultural and political ties to Djibouti and Somaliland. Local relationships enable data collection and community partnerships others cannot replicate.

Roadmap

Current stage and next validation.

June 22, 2026

SASR Eval: Public dataset validation

Validated on the skydheere/soomali-asr-dataset test split. Fine-tuned Whisper small (244M parameters) achieves mean WER 0.368.

August 2026

SASR Eval broadcast validation: Broadcast data validation

Validation on in-domain broadcast recordings. Tests generalization from public dataset to real regional media archives.

Data platform in development. We are building infrastructure to process and analyze decades of Horn of Africa broadcast and cultural archives. Detailed archive specifications are available to qualified partners under NDA.

Engagement

Actively engaging with cultural and academic partners.

Phoenix Rising Group is in active outreach with universities, cultural institutions, and language preservation organizations.

Universities Cultural Institutions Diaspora Organizations NLP Research Labs