README.md

metadata

title: README
emoji: 🎙️
colorFrom: green
colorTo: gray
sdk: static
pinned: false

Voice AI for Gulf Arabic: Saudi and Qatari.

Open models for Arabic as it is spoken on the phone, not as it is written in the news. Saudi dialects (Najdi, Hijazi, Khaleeji) are trained, measured and published today. Qatari is next.

🎙 Saudi STT 💬 Saudi Chat 4B 🗣 Saudi Speak 📚 Saudi Dialogues voho.ai →

Head to head · held-out tests

Beats OpenAI's speech model on Saudi Arabic.

Word errors on Saudi speech: 39.2% vs 103.9%. 62% fewer.

vs OpenAI Whisper Small · 4,582 held-out Saudi clips

Beats Google's Gemini at spoken Saudi, at a fraction of the size.

chrF++ to real Saudi speech: 75.7 vs 68.7, with a 0.6B model.

vs Gemini 2.5 Flash · 1,000 held-out Saudi sentences

Sounds more Saudi than Alibaba's Qwen, in a fifth of the words.

Replies that read as Saudi: 89.8% vs 62.3%. 6.4 words a reply vs 30.2.

vs Qwen3-4B-Instruct · 400 held-out questions

The stack: hear → decide → say

voho-saudi-stt-small CC BY-NC-SA

Saudi speech to text · 0.24B

voho-saudi-chat-4b Apache 2.0

Answers in spoken Najdi · 4B · GGUF

voho-saudi-speak-0.6b CC BY-NC-SA

Formal Arabic to spoken Saudi · 0.6B · GGUF · ONNX

voho-saudi-dialogues Apache 2.0

13,156 Najdi conversations across eight enterprise sectors

Quick facts

Gulf Arabic

Saudi today · Qatari next

Open models

3 + 1 dataset

Sizes

0.24B to 4B

Delivering

Aramco AI call centre

Run one now

ollama run hf.co/VohoAI/voho-saudi-chat-4b-GGUF:Q4_K_M

Then ask, in Najdi: أبي أحجز موعد بكرة الصبح، فيه وقت فاضي؟

The chat model and dataset are Apache 2.0, for commercial use. The speech models are CC BY-NC-SA 4.0; for production speech use the Voho API. · voho.ai/models