kanary-v2-1b-ko-34800
Korean controllable ASR checkpoint trained for 34,800 optimizer steps in two stages (8,700 untagged/exposure steps followed by 26,100 prompt-supervised steps). The model supports independent controls for number formatting, foreign word script, and punctuation.
This package is a release candidate. It is the best completed v2 checkpoint by the checkpoint-selection metric available on 2026-09-14: 7.94% CER on the KsponSpeech development view, compared with 7.98% for the previously selected v2 checkpoint. Evaluation of this exact checkpoint on KconfSpeech, KsponSpeech eval-clean, Zeroth, and FLEURS-ko was still queued when the package was created. Do not describe it as the best model on those test sets until the queued reports are added.
Checkpoint
| Item | Value |
|---|---|
| File | kanary-v2-1b-ko-34800.nemo |
| Size | 3,800,760,320 bytes |
| SHA-256 | cb12edf581adcb1561b3c7cc4edaae59d7a2ef0dc6557f5941daf153a3a2e0b8 |
| Parameters | 948,740,320 |
| Selection metric | KsponSpeech dev CER, 7.94% |
| Selected stage-2 step | 26,000; final export after step 26,100 |
| Total training budget | 8,700 + 26,100 = 34,800 steps |
The .nemo archive contains model weights, model configuration, and both
tokenizers. provenance.json records the exact local experiment identity and
checksums. Training configuration snapshots are included under training/.
Usage
The checkpoint declares the project-specific kanary prompt format. Install
this model repository so that kanary_prompt is importable before loading the
model:
pip install git+https://huggingface.co/lee1jun/kanary-v2-1b-ko-34800
import kanary_prompt
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained(
"lee1jun/kanary-v2-1b-ko-34800"
)
hyps = kanary_prompt.transcribe(
model,
["sample.wav"],
source_lang="ko",
target_lang="ko",
pnc="True",
itn="itn",
foreign="foreign:en",
)
print(hyps[0].text)
The extra slots must travel through a manifest; NeMo's ordinary
model.transcribe() interface does not forward the project-specific itn and
foreign fields. kanary_prompt.transcribe() creates that manifest.
Prompt controls
| Slot | Values | Intended behavior |
|---|---|---|
itn | itn, noitn, itn:undefined | digits/symbols, spoken-form Hangul, or model choice |
foreign | foreign:en, foreign:ko, foreign:undefined | Latin script, Hangul transcription, or model choice |
pnc | True, False | punctuation on or off |
Only Korean (ko) is supported for source_lang and target_lang.
Training
The model was initialized from the Korean-tokenizer Canary v2 graft and trained on AI Hub KsponSpeech, KtelSpeech, and KconfSpeech.
- Stage 1: 8,700 steps on the three-corpus exposure view.
- Stage 2: 26,100 steps on prompt-supervised ITN/foreign/PNC views, including non-applicable examples used to teach restraint.
- Optimizer: AdamW, learning rate
1e-4, inverse-square-root schedule, 1,000 warmup steps in each stage. - Batch duration: 1,800 seconds/GPU on four GPUs, bf16 mixed precision.
- Seed: 42 for the recovered stage-2 run. The original stage-1 random state was not replayed.
The source corpora total approximately 8,684 hours before derived prompt views. Derived rows are not additional unique audio.
Current evaluation status
The 7.94% development CER is suitable for checkpoint selection, not a claim of test-set superiority. Earlier v2 checkpoints showed strong formatting control, but those numbers are not copied here because they were measured from different weights. Update this section only from the reports for this exact SHA-256.
Limitations
- Korean only.
- A requested script marker is weaker evidence than exact recognition of the requested span.
- Rule-derived and LLM-generated multi-reference candidates require audio-based human approval before they can be treated as benchmark ground truth.
- The training data includes licensed AI Hub corpora; redistributing the model does not redistribute those datasets or supersede their terms.
License
CC-BY-4.0 for the checkpoint, following the existing Kanary release lineage. The source datasets retain their own terms.