rasr_sample

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - SV-SE dataset. It achieves the following results on the evaluation set:

• Loss: 0.3147
• Wer: 0.2676

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: 7.5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.3332	1.45	500	3.3031	1.0
2.9272	2.91	1000	2.9353	0.9970
2.0736	4.36	1500	1.1565	0.8714
1.7339	5.81	2000	0.7156	0.6688
1.5989	7.27	2500	0.5791	0.5519
1.4916	8.72	3000	0.5038	0.5169
1.4562	10.17	3500	0.4861	0.4805
1.3893	11.63	4000	0.4584	0.4761
1.3797	13.08	4500	0.4298	0.4686
1.3508	14.53	5000	0.4138	0.3744
1.3165	15.99	5500	0.4015	0.3578
1.281	17.44	6000	0.3883	0.3472
1.2682	18.89	6500	0.3904	0.3434
1.2477	20.35	7000	0.3726	0.3321
1.2364	21.8	7500	0.3685	0.3281
1.2041	23.26	8000	0.3597	0.3194
1.1901	24.71	8500	0.3542	0.3203
1.1903	26.16	9000	0.3500	0.3138
1.1677	27.61	9500	0.3458	0.3067
1.1718	29.07	10000	0.3595	0.3112
1.1562	30.52	10500	0.3433	0.3022
1.1392	31.97	11000	0.3440	0.2936
1.1258	33.43	11500	0.3396	0.2950
1.1067	34.88	12000	0.3379	0.2939
1.0953	36.34	12500	0.3370	0.2868
1.0835	37.79	13000	0.3317	0.2860
1.0772	39.24	13500	0.3302	0.2854
1.0853	40.7	14000	0.3265	0.2783
1.0689	42.15	14500	0.3306	0.2770
1.0394	43.6	15000	0.3233	0.2757
1.0581	45.06	15500	0.3199	0.2713
1.0362	46.51	16000	0.3154	0.2683
1.0406	47.96	16500	0.3176	0.2688
1.0082	49.42	17000	0.3149	0.2679

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0