OpenVoiceOS/whisper-large-v3-arabic-dialectal-v2-onnx

whisper-large-v3-arabic-dialectal-v2 — ONNX

ONNX export of oddadmix/whisper-large-v3-arabic-dialectal-v2 (Whisper large-v3 fine-tuned by oddadmix for dialectal Arabic, base model openai/whisper-large-v3) for onnx-asr (standard whisper model type — works with stock onnx-asr, no patches needed). fp32 and int8 variants included.

License: apache-2.0 (verbatim from the base model card).

Export note: this model's merged decoder graph exceeds the 2GB protobuf limit, so export used a patched optimum save path (external-data save bypassing the ByteSize() overflow in check_and_save_model). int8 was produced by quantizing the pre-merge decoders (decoder_model, decoder_with_past_model) and re-merging with strict=False — the merged graph's If-subgraphs defeat direct quantization of the merged decoder.

Dialect coverage

Trained on oddadmix/lahgtna-v3-small (dialect-balanced, undiacritized targets). Per the source model card, evaluated across 13 Arabic dialects (Gulf/Saudi, Iraqi, Egyptian, Syrian, Bahraini, Yemeni, Palestinian, Lebanese, Libyan, Tunisian, Algerian, Moroccan, Sudanese). Gulf/Saudi is consistently the strongest dialect; Maghrebi dialects (esp. Tunisian) are the weakest across the whole oddadmix v2 family. This is the largest and most accurate model in the family. This ONNX export does not change accuracy; it reproduces the source model's behavior.

Usage

import onnx_asr
model = onnx_asr.load_model("whisper", "path/to/this/repo")  # or quantization="int8"
print(model.recognize("audio_16khz.wav", language="ar"))

Verification

Verified against FLEURS ar_eg clips (fp32 and int8): both produce fluent, near-identical, high-fidelity undiacritized Arabic transcriptions; fp32 and int8 outputs match almost word-for-word.

RTF (CPU, homelab box, heavy contention during export batch): fp32 ~3-6, int8 ~2.2-3.6. Expect substantially better RTF under normal (uncontended) load.