Modelpublic

Reza2kn/canary-180m-persian-clean10-fast-asr

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

0likes48downloads

Model Card

Persian-heavy Canary 180M clean10 ASR checkpoints

Experimental full-parameter ASR-only fine-tunes of nvidia/canary-180m-flash with a Persian-heavy tokenizer.

Data mix: Persian clean10 AWQ-WER selector plus a small English FLEURS retention slice. Filtered train split: 16,273 rows / 10.416h; validation: 330 rows / 0.207h.

Artifacts:

—canary_180m_persian_clean10_gpu0_bd640.nemo: 2000 steps, batch_duration 640, final checkpoint.
—canary_180m_persian_clean10_gpu1_bd320.nemo: 2000 steps, batch_duration 320, final checkpoint.

Quick validation on the clean10 held-out split is included under eval/. This split is tiny and not a production benchmark; current results are poor and should be treated as a smoke/restore check:

—GPU0/bd640: WER 1.1273, CER 1.0211 over 330 rows.
—GPU1/bd320: WER 1.3428, CER 1.0248 over 330 rows.

These are research checkpoints, not production Android exports.