README.md

metadata

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 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