ctrlv-speechrecognition-model

This model is a fine-tuned version of facebook/wav2vec2-base on the TIMIT dataset. It achieves the following results on the evaluation set:

• Loss: 0.4730
• Wer: 0.3031

Test WER in TIMIT dataset

• Wer: 0.189

Google Colab Notebook

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: 8
• 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: 60
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.53	3.45	500	1.4021	0.9307
0.6077	6.9	1000	0.4255	0.4353
0.2331	10.34	1500	0.3887	0.3650
0.1436	13.79	2000	0.3579	0.3393
0.1021	17.24	2500	0.4447	0.3440
0.0797	20.69	3000	0.4041	0.3291
0.0657	24.14	3500	0.4262	0.3368
0.0525	27.59	4000	0.4937	0.3429
0.0454	31.03	4500	0.4449	0.3244
0.0373	34.48	5000	0.4363	0.3288
0.0321	37.93	5500	0.4519	0.3204
0.0288	41.38	6000	0.4440	0.3145
0.0259	44.83	6500	0.4691	0.3182
0.0203	48.28	7000	0.5062	0.3162
0.0171	51.72	7500	0.4762	0.3129
0.0166	55.17	8000	0.4772	0.3090
0.0147	58.62	8500	0.4730	0.3031

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.10.3