arun100/whisper-base-bn-3

Whisper Base Bengali

This model is a fine-tuned version of arun100/whisper-base-bn on the mozilla-foundation/common_voice_16_0 bn dataset. It achieves the following results on the evaluation set:

• Loss: 0.2151
• Wer: 29.9236

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: 1e-06
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 8500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2476	1.72	500	0.2751	36.1695
0.2284	3.43	1000	0.2622	35.1668
0.213	5.15	1500	0.2524	34.2022
0.2048	6.86	2000	0.2447	33.5266
0.1948	8.58	2500	0.2382	32.7495
0.1852	10.29	3000	0.2334	32.2322
0.1789	12.01	3500	0.2295	31.7244
0.1738	13.72	4000	0.2260	31.2341
0.166	15.44	4500	0.2236	30.9562
0.1629	17.15	5000	0.2214	30.8171
0.1636	18.87	5500	0.2194	30.4368
0.1578	20.58	6000	0.2181	30.2520
0.1628	22.3	6500	0.2170	30.1858
0.1566	24.01	7000	0.2161	30.0694
0.1564	25.73	7500	0.2156	29.9943
0.1545	27.44	8000	0.2153	29.9294
0.1548	29.16	8500	0.2151	29.9236

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.16.2.dev0
• Tokenizers 0.15.0