---
title: Nawah ASR 118M v5
emoji: 🎧
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
short_description: تفريغ صوتي عربي بموديل Emhotob 50M
---

# Nawah-ASR-118M-v5

Arabic ASR where a **50M Arabic LLM writes the transcript** -- and, as of v5, the acoustic
encoder is ours too: a 49.86M Whisper-architecture model trained from scratch on 866 h of
Egyptian Arabic, replacing whisper-small. **No OpenAI weights in the stack.**
[`oddadmix/50M-2048-Emhotob`](https://huggingface.co/oddadmix/50M-2048-Emhotob) does the generating.

Record or upload a clip (up to 30s). The demo reports the **audio token count**, which tracks the
clip's real duration at 50 tokens/sec -- a 3-second clip costs 150 tokens, not the 1500 a fixed
30-second window would spend on silence.

Scored on a 3,165-clip video-disjoint MASC split (same clips, same normalizer for every row):

| model | encoder | decoder | WER | CER |
|---|---|---|---|---|
| **`Nawah-ASR-118M-v5`** (this demo) | **ours, 49.86M from scratch** | Emhotob 50M | **0.3358** | **0.1420** |
| `Nawah-ASR-50M-v3` | whisper-small 88M | Emhotob 50M | 0.3501 | 0.1460 |
| `Nawah-ASR-118M-v4` (v5 before GRPO) | ours, 49.86M | Emhotob 50M | 0.3682 | 0.1571 |
| `Nawah-ASR-50M-v2` | whisper-small 88M | Emhotob 50M | 0.3614 | 0.1549 |
| `Nawah-ASR-50M-v1.1` (300 h) | whisper-small 88M | Emhotob 50M | 0.4109 | 0.1815 |
| `Nawah-ASR-50M-v1` (60 h) | whisper-small 88M | Emhotob 50M | 0.6145 | 0.2936 |
| `openai/whisper-small` (zero-shot) | -- | Whisper 153M | 0.5739 | 0.2062 |
| `whisper-small-arabic-dialectal` | -- | Whisper 153M | 0.6775 | 0.2244 |

v5 = v4 + a GRPO phase (sequence-level RL against the dataset's own labels), worth -0.032 WER.
It is the first model here to beat the whisper-small-encoder versions while being smaller (118M
vs ~138M) and ~50% faster to decode.

With the encoder fixed, swapping Whisper's 153M decoder for Emhotob's 50M takes WER 0.5739 ->
0.4109. Caveat: this is MASC's home turf and both Whisper checkpoints are out-of-domain here --
the tell is that the dialect-tuned one scores *worse* than the base model on this MSA-leaning data.

Model: [`oddadmix/Nawah-ASR-50M-v1.1`](https://huggingface.co/oddadmix/Nawah-ASR-50M-v1.1)