wav2vec2-base-repro-timit

This model is a fine-tuned version of patrickvonplaten/wav2vec2-base-repro-960h-libri-85k-steps on the TIMIT_ASR - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.8562
• Wer: 0.5484

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
5.9793	0.69	100	5.4532	1.0
2.9066	1.38	200	2.9070	1.0
2.2562	2.07	300	2.0323	1.0
1.5273	2.76	400	1.1510	0.8001
1.1085	3.45	500	0.9521	0.7053
0.813	4.14	600	0.8617	0.6702
0.8434	4.83	700	0.8068	0.6393
0.9631	5.52	800	0.7863	0.6248
0.707	6.21	900	0.7476	0.5973
0.5568	6.9	1000	0.7350	0.5911
0.6171	7.59	1100	0.7171	0.5841
0.7011	8.28	1200	0.7318	0.5798
0.5546	8.97	1300	0.7447	0.5767
0.4278	9.66	1400	0.7481	0.5650
0.3576	10.34	1500	0.7443	0.5713
0.5506	11.03	1600	0.7574	0.5664
0.4127	11.72	1700	0.8043	0.5631
0.3251	12.41	1800	0.7738	0.5550
0.3119	13.1	1900	0.7829	0.5516
0.4371	13.79	2000	0.8025	0.5556
0.3772	14.48	2100	0.8451	0.5559
0.2942	15.17	2200	0.8300	0.5556
0.2503	15.86	2300	0.8417	0.5541
0.3671	16.55	2400	0.8568	0.5528
0.3867	17.24	2500	0.8521	0.5510
0.2614	17.93	2600	0.8479	0.5523
0.2441	18.62	2700	0.8558	0.5494
0.3059	19.31	2800	0.8553	0.5474
0.3734	20.0	2900	0.8562	0.5484

Framework versions

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