unispeech-sat-large-timit-ft

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

• Loss: 0.6074
• Wer: 0.3880

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: 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
6.2516	0.69	100	5.8638	1.0
2.9596	1.38	200	2.9550	1.0
2.8831	2.07	300	2.8547	1.0
2.3223	2.76	400	2.2044	1.0063
1.2104	3.45	500	1.0845	0.7706
0.6779	4.14	600	0.7342	0.5663
0.6319	4.83	700	0.6054	0.4881
0.664	5.52	800	0.5808	0.4913
0.402	6.21	900	0.5647	0.4611
0.3176	6.9	1000	0.5211	0.4440
0.3392	7.59	1100	0.5187	0.4359
0.3888	8.28	1200	0.5501	0.4391
0.2874	8.97	1300	0.5249	0.4148
0.208	9.66	1400	0.5407	0.4152
0.1457	10.34	1500	0.5722	0.4155
0.2375	11.03	1600	0.5780	0.4059
0.2111	11.72	1700	0.5823	0.4094
0.1422	12.41	1800	0.5754	0.3977
0.125	13.1	1900	0.5784	0.4031
0.1996	13.79	2000	0.5630	0.3956
0.1747	14.48	2100	0.5880	0.3964
0.1263	15.17	2200	0.5987	0.3951
0.11	15.86	2300	0.5688	0.3964
0.1411	16.55	2400	0.6223	0.3906
0.1647	17.24	2500	0.6135	0.3960
0.1162	17.93	2600	0.6224	0.3960
0.098	18.62	2700	0.6017	0.3907
0.1183	19.31	2800	0.6121	0.3885
0.1717	20.0	2900	0.6074	0.3880

Framework versions

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