qwen3-asr-0.6b-arabic-dialectal-v2-onnx

qwen3-asr-0.6b-arabic-dialectal-v2 — ONNX

ONNX export of oddadmix/qwen3-asr-0.6b-arabic-dialectal-v2 (fine-tuned by oddadmix for dialectal Arabic, base model Qwen/Qwen3-ASR-0.6B-hf) for onnx-asr's speech-llm model type (three-graph contract: encoder + embed_tokens + decoder with KV cache). 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), the same corpus as the other oddadmix v2 dialectal releases. Per the family's pattern, expect strongest results on Gulf/Saudi and Egyptian, weakest on Maghrebi dialects (Tunisian, Algerian, Moroccan). This ONNX export does not change accuracy; it reproduces the source model's behavior.

Usage

import onnx_asr
model = onnx_asr.load_model("speech-llm", "path/to/this/repo")  # or quantization="int8"
print(model.recognize("audio_16khz.wav"))

Requires an onnx-asr build with speech-llm support (the Qwen3-ASR three-graph contract landed in TigreGotico's onnx-asr fork; verify your onnx-asr install includes onnx_asr/models/speech_llm.py before loading this model type).

Verification

Verified against FLEURS ar_eg clips, comparing against the native transformers Qwen3ASRForConditionalGeneration output on the same fine-tuned checkpoint: fp32 ONNX output is character-exact vs. native on 4/5 clips (one clip differs by a single word-boundary artifact); int8 shows the expected minor quality drop from dynamic weight quantization of the 0.6B decoder but stays intelligible and dialect-faithful.

RTF (CPU, homelab box, shared load): fp32 ~0.4-0.6, int8 ~0.2-0.5.