Model Card for whisper-tiny-ko-vietnamese-accent

Model Description

This model is a fine-tuned version of openai/whisper-tiny for Korean automatic speech recognition (ASR), specifically adapted to Korean speech produced by native Vietnamese speakers. It was developed as part of a research project on improving ASR fairness and inclusiveness for non-native Korean speakers.

Training Data

• Source: AI Hub "AI Training Data for Foreign Korean Speech" (Vietnamese, Validation subset)
• Train set: 1,366 utterances
• Validation set: 341 utterances
• Speakers do not overlap between train and validation sets.
• Only "Reading" (scripted prompt-reading) utterances were used; free-response utterances were excluded.
• Ground truth transcripts: ReadingLabelText column.

Training Procedure

• Base model: openai/whisper-tiny
• Steps: 500 (batch size 8, learning rate 1e-5, warmup steps 50)
• Checkpoint selection: best checkpoint selected by lowest Validation WER (load_best_model_at_end=True, metric_for_best_model="wer")
• Evaluated using direct WhisperProcessor + WhisperForConditionalGeneration inference (not the pipeline() API) for consistency between baseline and fine-tuned evaluation.

Evaluation Results

	WER	CER
Baseline / Validation	58.27%	34.97%
Fine-tuned / Validation	18.21%	9.33%
Baseline / Test*	52.73%	22.49%
Fine-tuned / Test*	44.27%	21.06%

*Test set consists of 77 utterances from a pizza/chicken phone-order scenario, entirely held out from training (no overlapping sentences or speakers with Train/Validation).

Intended Use

Research on Korean ASR fairness for non-native speakers. Not intended for production deployment without further evaluation.

Limitations

• Trained on a relatively small dataset (1,366 utterances) from a single language background (Vietnamese).
• Performance improvement is larger on in-domain validation data than on the held-out test scenario, suggesting limited generalization to entirely new contexts.