sat-base

This model is a fine-tuned version of microsoft/unispeech-sat-base on the TIMIT_ASR - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.7014
• Wer: 0.5374

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: 3e-05
• train_batch_size: 32
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.9958	0.69	100	6.7171	1.0
3.0453	1.38	200	3.0374	1.0
2.9989	2.07	300	2.9807	1.0
2.969	2.76	400	2.9579	1.0
2.903	3.45	500	2.9072	1.0
2.8565	4.14	600	2.8804	1.0
2.8195	4.83	700	2.7916	1.0
2.3134	5.52	800	2.1456	1.0004
1.5475	6.21	900	1.4663	0.9549
1.1295	6.9	1000	1.1140	0.7227
1.0181	7.59	1100	0.9258	0.6497
1.0252	8.28	1200	0.8430	0.6255
0.835	8.97	1300	0.8063	0.6032
0.662	9.66	1400	0.7595	0.5931
0.5558	10.34	1500	0.7322	0.5819
0.7596	11.03	1600	0.7120	0.5708
0.6169	11.72	1700	0.7073	0.5606
0.4565	12.41	1800	0.7124	0.5586
0.4554	13.1	1900	0.6880	0.5501
0.6216	13.79	2000	0.6783	0.5494
0.5393	14.48	2100	0.7067	0.5499
0.4095	15.17	2200	0.7014	0.5438
0.3551	15.86	2300	0.7000	0.5426
0.5112	16.55	2400	0.6866	0.5426
0.5139	17.24	2500	0.7134	0.5446
0.3638	17.93	2600	0.7130	0.5434
0.3327	18.62	2700	0.6980	0.5377
0.4385	19.31	2800	0.7017	0.5390
0.4986	20.0	2900	0.7014	0.5374

Framework versions

• Transformers 4.12.0.dev0
• Pytorch 1.8.1
• Datasets 1.14.1.dev0
• Tokenizers 0.10.3