whisper-medium-ko-1195h

This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1552
• Wer: 8.6411 (is improved against 10.4449 from jangmin/whisper-small-ko-1159h)

Model description

The model was trained to transcript the audio sources into Korean text.

Intended uses & limitations

More information needed

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". I intentionally gathered 796 hours of audio from the first dataset and 363 hours of audio from the second dataset (This includes statistics for the training data only, and excludes information about the validation data.).

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 10
• eval_batch_size: 10
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 59151
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0782	0.33	6572	0.1833	10.9268
0.07	0.67	13144	0.1680	10.3611
0.0605	1.0	19716	0.1600	9.9357
0.0345	1.33	26288	0.1573	9.4492
0.0365	1.67	32860	0.1518	9.3395
0.0339	2.0	39432	0.1478	8.9811
0.0176	2.33	46004	0.1596	9.1702
0.0159	2.67	52576	0.1572	8.6746
0.0141	3.0	59148	0.1552	8.6411

Framework versions

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