Reza2kn/canary-180m-persian-semiclean31-staged-smart-init

Persian-heavy Canary 180M staged smart-init ASR

Experimental ASR-only adaptation of nvidia/canary-180m-flash for Persian-first bilingual ASR.

This checkpoint uses the newer staged adaptation path:

• —Fresh Persian-heavy bilingual SentencePiece tokenizer.
• —Smart vocabulary/embedding initialization from the original Canary tokenizer where possible.
• —Stage 1: decoder/head adaptation with the encoder frozen.
• —Stage 2: full-parameter continuation with a short initial encoder freeze.

Data mix:

• —Persian: Reza2kn/persian-asr-semi-clean-31h-awq-wer selected/cleaned audio+text only.
• —English: small FLEURS retention slice.
• —Train split: 46,006 rows, about 31.742 hours.
• —Validation split: 938 rows, about 0.652 hours.

Validation on the internal portable held-out split:

• —Rows: 938
• —WER: 0.341208 (34.12%)
• —CER: 0.195946 (19.59%)

Artifact:

• —canary_180m_persian_semiclean31_staged_smart_gpu1_bd700.nemo
• —SHA256: 77fe2c46c30a507440b7129bb2efbb8e9b0e18622346509c7c46e99af16adb49

This is still a research checkpoint, not yet an Android/CoreML/ONNX export. The earlier non-smart-init semiclean31 checkpoint was much worse, around 102% WER; this staged smart-init checkpoint is the first run where the adaptation is clearly learning.