whisper-medium-ko-normalized-1273h

This model is a fine-tuned version of openai/whisper-medium on a custom dataset for improving Korean speech recognition. It achieves the following results on the evaluation set:

• Loss: 0.1254
• Wer: 0.0551

Model description

The model was a fine-tuned version of openai/whisper-medium transcript the Korean audio sources into text. It was trained on GCP's a2-highgpu-1g (a100-40G) for 26 hours with about $90.

Intended uses & limitations

This model was trained to extend the performance of the original whisper model for Korean transcription task.

Training and evaluation data

I downloaded all data from AI-HUB (https://aihub.or.kr/). Two datasets, in particular, caught my attention: "Instruction Audio Set" and "Noisy Conversation Audio Set". Following indicates the hours information for each dastset.

dataset name	train_split (hours)	validation_split (hours)
Instruction Audio Set	910	105
Noisy Conversation Audio Set	363	76

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 24
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0588	1.0	8775	0.1225	0.0604
0.0287	2.0	17550	0.1186	0.0567
0.0148	3.0	26325	0.1254	0.0551

Framework versions

• Transformers 4.28.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.11.0
• Tokenizers 0.13.2

Evaluation Result for the dataset google/fleurs

The trained model is evaluated on the test split of subset ko_kr from the dataset google/fleurs. Please note that the model was not trained on the train split from the dataset.

model	Wer
openai/whisper	0.2469
this model	0.2189