Organisation card: the stack, the data, the method

README.md

@@ -1,10 +1,61 @@
+# Voho
+
+**Voice AI that speaks Saudi.** Models trained by Voho, for the phone line — evaluated on held-out data we didn't write, numbers published, before and after.
+
+Most Arabic AI is built for Modern Standard Arabic and tested on it. Nobody in Riyadh talks that way on a call. Every model here exists because a general model failed a specific Saudi task, and each card shows by how much.
+
+---
+
+## The stack
+
+Three models, one pipeline: hear → decide → say.
+
+| | Model | Does | Measured |
+| --- | --- | --- | --- |
+| 🎙 | [**voho-saudi-stt-small**](https://huggingface.co/VohoAI/voho-saudi-stt-small) | Saudi speech → text | WER **104% → 39%**, CER **69% → 17%** on 4,582 Saudi test clips, per dialect |
+| 💬 | [**voho-saudi-chat-4b**](https://huggingface.co/VohoAI/voho-saudi-chat-4b) | Answers in spoken Najdi | Gulf **62% → 90%** of replies (reference 94.5%), MSA 0.8%, replies **30 → 6 words** |
+| 🗣 | [**voho-saudi-speak-0.6b**](https://huggingface.co/VohoAI/voho-saudi-speak-0.6b) | Formal Arabic → how a Saudi says it | chrF++ **52 → 76** vs real Saudi speech, across Najdi, Hijazi, Khaleeji |
+
+Every "before" is the untrained base model on the same test set. Every test set is held out and, for the chat model, third-party. Each card has the full table, the examples, and the limitations.
+
+---
+
+## The data
+
+[**voho-saudi-dialogues**](https://huggingface.co/datasets/VohoAI/voho-saudi-dialogues) — 13,156 multi-turn Najdi conversations, 105,808 turns. Apache 2.0.
+
+7,603 service calls across oil and gas, utilities, telecom, banking, government, healthcare, logistics and facilities — a rig shift handover, a permit-to-work request, a SADAD dispute. Nothing like it existed in the open. Plus 5,553 everyday conversations, because a model trained only on service calls treats every sentence as a ticket.
+
+Every dialogue passed a Najdi lexicon gate and an independent dialect classifier before it was kept.
+
+---
+
+## How we build
+
+- **Own data for the job.** When the open data covers `daily_life` and `travel` and the customer needs a refinery, we build the refinery data and publish it.
+- **Gated, not asserted.** Generated text is filtered by a classifier that had no part in training. The keep rate is on the card.
+- **Baselined.** Untrained base and reference ceiling on every card, so a number means something.
+- **Evaluated on what we didn't write.** The chat model's test set is a published third-party corpus it never trained on.
+- **Small and cheap.** 0.6B to 4B, one L4, a few dollars a run. Bigger was never the bottleneck; data was.
+
+The method is documented on each model card and reusable across languages — this is how Voho trains, not a one-off.
+
+---
+
+## Run one now
+
+```sh
+ollama run hf.co/VohoAI/voho-saudi-chat-4b-GGUF:Q4_K_M
+```
+
+Then, in Najdi: أبي أحجز موعد بكرة الصبح، فيه وقت فاضي؟
+
 ---
-title: README
-emoji: 🚀
-colorFrom: green
-colorTo: indigo
-sdk: static
-pinned: false
+
+## Licences
+
+The chat model and the dataset are **Apache 2.0** — use them commercially. The two speech models are **CC BY-NC-SA 4.0**; for production speech, use the [Voho API](https://docs.voho.ai).
+
 ---
 
-Edit this `README.md` markdown file to author your organization card.
+**[voho.ai](https://voho.ai)** · **[docs.voho.ai](https://docs.voho.ai)** · Saudi Arabic voice agents for enterprise