p4b/whisper-small-ko-fl

Whisper Small Ko(FLUERS) - by p4b

This model is a fine-tuned version of openai/whisper-small on the FLUERS Korean dataset. It achieves the following results on the evaluation set:

• Loss: 0.2893
• Wer: 19.2

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

Dataset filtering

Some of datas from FLUERS are not used for training and evaluation. Most of filtered datas are not fit to model or including non-korean symbols.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• train_batch_size: 96
• eval_batch_size: 64
• seed: 42
• distributed_type: multi-GPU
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3016	32.0	800	0.4048	140.4726
0.0451	64.0	1600	0.2893	19.2043
0.0169	96.0	2400	0.3110	20.2513
0.0092	128.0	3200	0.3240	20.0419
0.0062	160.0	4000	0.3335	20.0419
0.0045	192.0	4800	0.3416	20.0718
0.0035	224.0	5600	0.3501	20.1615
0.0028	256.0	6400	0.3562	20.3709
0.0024	288.0	7200	0.3618	20.0120
0.002	320.0	8000	0.3669	20.1017
0.0017	352.0	8800	0.3704	20.1914
0.0017	384.0	9600	0.3723	20.2513

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.14.0.dev20221208+cu116
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2