ko-phone-int8

An ONNX, int8-quantized copy of slplab/wav2vec2-xls-r-300m_phone-mfa_korean, a Korean phone (MFA phoneset) recognition model. It is used by CheGROW Studio Pro to find repeated or added reading in narration recordings.

Original work

• Model: slplab/wav2vec2-xls-r-300m_phone-mfa_korean (model management by excalibur12), licensed under Apache-2.0. It is a fine-tuned version of facebook/wav2vec2-xls-r-300m.
• Training data, as described by the original authors: Phonetically Balanced Native Korean Read-speech Corpus. See the original model card for details. The original card does not state the terms of the training data.

Changes made

• Converted the PyTorch weights (revision e26ff9dfb62169acf445d0060ef56863c018b20e) to ONNX (opset 17).
• Dynamically quantized to int8 with ONNX Runtime (per-channel, MatMul only): onnx/model_quantized.onnx.
• config.json, preprocessor_config.json, vocab.json, tokenizer_config.json, special_tokens_map.json and added_tokens.json are unchanged copies of the original files.

Quantization changes the numerical outputs slightly compared with the original model.

Usage (Transformers.js)

import { AutoModelForCTC } from '@huggingface/transformers';
const model = await AutoModelForCTC.from_pretrained('chegrowk/ko-phone-int8', { dtype: 'q8' });
// input_values: 16 kHz mono audio, normalized to zero mean and unit variance

Citation

If you use this model, please cite the original authors:

Yeo, E.*, Ryu, H.*, Lee, J., Kim, S., & Chung, M. (2023). Comparison of L2 Korean pronunciation error patterns from five L1 backgrounds by using automatic phonetic transcription. Proceedings of the 20th International Congress of Phonetic Sciences, pp. 2696–2700.

License

Apache License 2.0 — see LICENSE. Provided "as is", without warranties of any kind.