ymoslem/whisper-base-ga2en-v2.1

Whisper Base GA-EN Speech Translation

This model is a fine-tuned version of openai/whisper-base on an unknown dataset. The best model (this version) is at checkpoint 1800, epoch 1.94, and it achieves the following results on the evaluation set:

• Loss: 1.6780
• Bleu: 22.52
• Chrf: 39.24
• Wer: 76.7222

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: 32
• 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: 0.03
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Bleu	Chrf	Validation Loss	Wer
2.4277	0.11	100	4.21	19.8	2.2931	144.9797
2.1912	0.22	200	6.79	22.82	2.0213	152.5439
1.9199	0.32	300	6.62	24.85	1.9041	180.0090
1.7525	0.43	400	13.14	30.42	1.8026	98.1090
1.6623	0.54	500	17.73	34.37	1.7467	90.5448
1.4937	0.65	600	16.85	33.97	1.7301	92.3458
1.3587	0.76	700	14.77	33.3	1.6499	101.7109
1.274	0.86	800	18.28	35.46	1.6641	89.1941
1.1514	0.97	900	21.17	37.05	1.6172	80.1441
0.6932	1.08	1000	16.81	35.35	1.6421	99.0095
0.8294	1.19	1100	1.6699	18.49	36.78	90.9500
0.7662	1.29	1200	1.7147	17.15	34.75	95.8577
0.7704	1.4	1300	1.6752	15.65	35.08	104.2774
0.7333	1.51	1400	1.6812	19.17	36.87	89.2841
0.6879	1.62	1500	1.6719	19.09	37.98	84.6015
0.6297	1.73	1600	1.6847	19.43	37.28	89.5092
0.5843	1.83	1700	1.6659	17.74	38.08	98.1990
0.5342	1.94	1800	1.6780	22.52	39.24	76.7222
0.2743	2.05	1900	1.7151	22.48	39.05	78.8834
0.2932	2.16	2000	1.7044	17.65	38.01	99.2796

Framework versions

• Transformers 4.39.3
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2