whisper-tiny-hi

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

• Loss: 2.2091
• Wer: 86.2832

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• training_steps: 500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.5332	3.57	25	1.7565	89.3805
0.3504	7.14	50	1.6160	90.2655
0.0551	10.71	75	1.8524	93.3628
0.0285	14.29	100	1.9261	123.0088
0.022	17.86	125	2.0688	92.9204
0.0075	21.43	150	2.0535	89.3805
0.0054	25.0	175	2.1533	86.7257
0.0016	28.57	200	2.1682	91.1504
0.001	32.14	225	2.2014	87.6106
0.001	35.71	250	2.1406	87.1681
0.0037	39.29	275	2.1968	88.0531
0.0012	42.86	300	2.1761	107.0796
0.0004	46.43	325	2.1874	88.0531
0.0003	50.0	350	2.2005	87.1681
0.0003	53.57	375	2.2018	87.1681
0.0002	57.14	400	2.2041	87.1681
0.0002	60.71	425	2.2055	86.7257
0.0002	64.29	450	2.2072	86.2832
0.0002	67.86	475	2.2089	87.1681
0.0002	71.43	500	2.2091	86.2832

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.1.0+cu121
• Datasets 2.15.0
• Tokenizers 0.15.0