ko-jamo-base (ONNX)

An ONNX export of Kkonjeong/wav2vec2-base-korean, used by CheGROW Studio Pro to find repeated or added words when checking a narration against its script.

No weights were changed. The model was exported as-is (float32, ONNX opset 17). On a test input, the ONNX output matched the original PyTorch model (largest difference 0.000009, same predicted symbol on every frame).

Files

• onnx/model.onnx — the model (input input_values, 16 kHz mono, mean 0 / variance 1; output logits)
• vocab.json and the tokenizer files — the original model's vocabulary (Korean jamo characters)
• config.json — the original model's configuration

Credits and licenses

• Original model: Kkonjeong/wav2vec2-base-korean by Jeonghyeon Park and Jaeyoung Kim. License: Apache License 2.0, as stated in the original model card.
• Base model: facebook/wav2vec2-base. License: Apache License 2.0.
• Training data of the original model: Zeroth-Korean (OpenSLR 40) by Lucas Jo (Atlas Guide Inc.) and Wonkyum Lee (Gridspace Inc.). License: CC BY 4.0.

This export is distributed under the Apache License 2.0 (see LICENSE).