Ku bood nuxurka

Wax kasta oo aan ururinno, si furan

Waxaa la cusboonaysiiyaa Isniin kasta. Haddii tiro hoos u dhacdo, taasna waad arki doontaa.

La cusboonaysiiyay 28 Sep 2026 · xog tusaale ah
Saacadaha la ansixiyay
47.6
+9.5 toddobaadkan
Ka qayb qaatayaasha firfircoon
612
+88 toddobaadkan
Clips la aqbalay
81%
+1 dhibichadaf: 80% bisha 6
Heerka khaladka (WER)
36%
−3 dhibcood, khalad yar58% baseline-ka

Saacadaha la ansixiyay si dhakhso ah ayay u korayaan

Wadarta saacadaha cod ee la ansixiyay, toddobaad kasta. Hadaf: 100 saacadood bisha 12.

01020304050T1T2T3T4T5T6T7T8T9T10T11T1247.6 hGoal 100 hours

Kuwa aan maqlayno

Qaybta saacadaha la ansixiyay ee lahjad kasta. Maay waa meesha ugu dayacan.

Maxaa-tiri, koonfur iyo Banaadir41%
Maxaa-tiri, waqooyi38%
Qurbajoog iyo Soomaali-Ingiriisi12%
Maay9%

Hadaf 20%. Duubista Maay-gu waa dhibco labanlaab.

Dumar46%
Rag54%

Version kasta oo model-ka ah khalad yar ayuu sameeyaa

Word error rate-ka tijaabada Soomaaliga ah. Inta hoose ayaa fiican.

20%30%40%50%60%Baselinev0.1v0.2v0.3v0.4v0.5v0.6Whisper + MMS baseline 58%36%

Kaydka Data Dividend

20% dakhliga license-ka iyo annotation-ka.

$1,240

Hal license (Sep 2026). Lacag bixinta xigta waa marka rubuca xirmo 31 Dec.

Lacag bixinKaydDad
Q4 2026 (sugaya)$1,240612
Q3 2026$0—

Waxa la daabacay

TaariikhDaabacaadWaxa isbeddelayLicense
22 Sep 2026Xuddun Speech v0.140 saacadood, 480 qof, 6 gobol, data card-kii koowaad30% furan (CC-BY 4.0), inta kale ganacsi
08 Sep 2026Somali ASR v0.6WER 39% → 36% kadib clips qurbajoogAPI (beta)
18 Aug 2026Somali test set1,000 jumladood, 3 Khabiir ayaa hubiyayCC-BY 4.0