Research
How each Voho model was built and measured.
14 September 2026 · Voho Saudi STT SmallHearing Saudi Arabic as it is spokenA 0.24B speech model trained on about 187,000 Saudi clips cuts word errors from 103.9% to 39.2% against the OpenAI model it started from, across Najdi, Hijazi and Khaleeji.Read the research →14 September 2026 · Voho Saudi Speak 0.6BSaying it the way Saudis say itA 0.6B model rewrites formal Arabic into Najdi, Hijazi or Khaleeji. On 1,000 held-out sentences it scores 75.7 chrF++ against what Saudi speakers actually said, above Gemini 2.5 Flash at 68.7.Read the research →19 September 2026 · Voho Saudi Chat 4BAnswering in Najdi, at phone-call lengthA 4B assistant whose replies read as Gulf 89.8% of the time, against 62.3% for the model it started from, at 6.4 words a turn. Apache 2.0, the dialogue data included.Read the research →