whisper-small-arabic-dialectal-v2-onnx

whisper-small-arabic-dialectal-v2 — ONNX

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

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

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). Reported overall WER 0.403 / CER 0.153 on the balanced test set; Gulf/Saudi is the strongest dialect (WER 0.227), Maghrebi dialects (esp. Tunisian, WER 0.606) are the weakest — consistent with the source card's own analysis. 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 undiacritized Arabic transcriptions consistent with the source model's reported accuracy (minor dialectal spelling variation vs. the diacritized FLEURS reference is expected and matches the source model's own behavior, not an export artifact). fp32 and int8 outputs match closely (int8 = dynamic weight-only quantization of the encoder + merged decoder).

RTF (CPU, AMD EPYC-class homelab box, shared load): fp32 ~0.2-0.3, int8 ~0.15-0.23.