Moodalka
What it is
whisper-somali-v5 is a Somali speech-to-text model built by Impetik. It is a fine-tune of openai/whisper-large-v3-turbo, an 809 million parameter model, trained on about 240 hours of curated Somali speech. It outputs Somali text with punctuation and capital letters.
To our knowledge it is the most accurate Somali speech recognition model available as of September 2026. It was built in five generations over three days, with an audited benchmark decontamination and checksum-verified artifacts. The full story is in our launch post.
Benchmarks
WER (word error rate) is the percentage of words the model got wrong. CER (character error rate) is the same measured per character. Lower is better.
| Test set | WER | CER |
|---|---|---|
| ANV conversational referee (845 clips, held-out shards, never trained on) | 22.3% | 7.0% |
| thiomi-5k test (read speech; train split was in training data) | 18.9% | 6.0% |
| 25 real-world phone-mic recordings (owner-verified transcripts) | 37.9% | 9.9% |
| BBC Somali broadcast sample | 3.7% | 0.7% |
On the same sets the stock facebook/mms-1b-all model scores about 44 to 47% WER. The best public Somali fine-tunes score 45 to 52% on the real-world set. Accuracy depends heavily on the audio. Broadcast and read speech is near-perfect. Noisy spontaneous phone audio is hardest.
Training data
| Source | Hours | License |
|---|---|---|
| Digital-Divide-Data/Somali-ASR-Subset-68H (Maxaa Tiri, reviewed) | ~150 | CC-BY-4.0 |
| African Next Voices Kenya (conversational; shards ending in 9 held out for evaluation) | 110 | CC-BY-4.0 |
| google/fleurs so_so train | ~10 | CC-BY-4.0 |
| thiomi/thiomi-5k som_Latn train | 8.7 | not published |
Attribution: Maseno University / African Next Voices, Digital Divide Data, Google FLEURS, and the thiomi-5k authors.
Clips were filtered to 1 to 30 seconds with non-empty transcripts and no digits, then packed into about 28-second composites. Training ran for 3 epochs on two A100 GPUs.
License
The base model is MIT licensed. Three of the four data sources are CC-BY-4.0. The thiomi-5k dataset has no published license, so the redistribution and commercial status of this model is unclear until that is resolved. We therefore treat v5 as a research preview. A commercially clean version, same pipeline without thiomi, can be produced on request.
Limitations
- Trained mostly on Somali collected in Kenya. Coverage of Somalia and diaspora dialects is untested beyond the owner’s samples and one BBC clip.
- English words inside Somali speech are often mangled. The training data is pure Somali.
- On non-speech or heavily degraded audio it may output fluent invented text rather than nothing.
- It may stop early at long pauses, and 30-second chunking can drop words at chunk boundaries.
- Numerals were filtered from training transcripts. Spoken numbers transcribe as words.
- Evaluation sets are modest, so the WER figures carry a few points of noise.
Contribute your voice
Record a clip on the home page, correct what the model heard and submit it. Every corrected clip goes into a Somali speech dataset that we will publish on Hugging Face under CC-BY-4.0, free for anyone to use.