ThejundiAI/dawwar-speech-models

Dawwar speech models (ONNX)

نماذج التفريغ النصي التي ينزّلها تطبيق دوّار من إعداداته، لتحويل الكلام إلى نص على جهاز المستخدم دون إنترنت.

These are ONNX conversions of open speech-recognition models, prepared for Dawwar, a free Windows app by ASLS AI. The app downloads them from its Settings and runs them entirely on the user's PC; no audio is sent anywhere. The app checks every file against SHA-256 hashes in a list signed with its release key, so only these exact files are accepted.

Folder	Model	What changed	Licence
cohere-ar/	Cohere Transcribe Arabic 07-2026 by Cohere Labs	Exported to ONNX (encoder, and a decoder with an explicit key/value cache). cpu/: int8 dynamic quantization. gpu/: fp16 encoder for DirectML.	Apache-2.0 (cohere-ar/LICENSE)
nvidia-ar/	NVIDIA FastConformer Hybrid Large, Arabic (pc) v1.0	CTC branch exported to ONNX with NeMo; MatMul weights quantized to int8.	CC BY 4.0
nvidia-en/	NVIDIA FastConformer Hybrid Large, English (pc)	CTC branch exported to ONNX with NeMo; MatMul weights quantized to int8.	CC BY 4.0
runtime/x64/	ONNX Runtime 1.24.4 (DirectML build), unmodified	—	MIT (runtime/x64/LICENSE, ThirdPartyNotices.txt)

Each model folder also holds mel.f32 (the model's mel filter bank) and vocab.json (its token list), extracted from the original model.

DirectML itself is not hosted here: the app uses the copy built into Windows, or fetches Microsoft's own Microsoft.AI.DirectML package from nuget.org on older Windows.

Accuracy notes from our tests: the Cohere conversion reproduces the original model's output word for word when given the same audio windows. The NVIDIA conversions match NeMo's own output; their int8 versions differ from fp32 in under 0.5% of characters on our samples. The NVIDIA Arabic model is trained mainly on Modern Standard Arabic and is weaker on dialects.

These files are provided as they are, without warranty. Cohere Labs and NVIDIA did not make, review or endorse these conversions or the Dawwar app. See LICENSES.md for the attributions.