wav2vec2-base-960h-librispeech-model

This model is a fine-tuned version of facebook/wav2vec2-base-960h on the LIBRI10H - ENG dataset. It achieves the following results on the evaluation set:

• Loss: 0.6832
• Wer: 0.5184

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: 0.003
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.477	1.1565	200	2.8801	1.0
2.8001	2.3130	400	2.1488	0.9999
1.6185	3.4696	600	1.2806	0.9032
1.359	4.6261	800	1.1836	0.8888
1.2731	5.7826	1000	1.1609	0.8760
1.2092	6.9391	1200	1.1091	0.8629
1.1681	8.0928	1400	1.0761	0.8544
1.1297	9.2493	1600	1.0792	0.8494
1.0982	10.4058	1800	1.0455	0.8353
1.073	11.5623	2000	1.0372	0.8361
1.0436	12.7188	2200	1.0217	0.8368
1.021	13.8754	2400	0.9912	0.8129
0.9893	15.0290	2600	0.9936	0.8073
0.964	16.1855	2800	0.9619	0.7934
0.9391	17.3420	3000	0.9557	0.7898
0.9159	18.4986	3200	0.9378	0.7797
0.8927	19.6551	3400	0.9074	0.7680
0.8714	20.8116	3600	0.9022	0.7623
0.8412	21.9681	3800	0.8682	0.7360
0.8083	23.1217	4000	0.8436	0.7153
0.7838	24.2783	4200	0.8584	0.7048
0.7533	25.4348	4400	0.8101	0.6912
0.7286	26.5913	4600	0.7933	0.6707
0.6965	27.7478	4800	0.8056	0.6681
0.6859	28.9043	5000	0.7554	0.6417
0.6529	30.0580	5200	0.7624	0.6291
0.6315	31.2145	5400	0.7508	0.6123
0.6144	32.3710	5600	0.7255	0.6056
0.5933	33.5275	5800	0.7546	0.6045
0.5827	34.6841	6000	0.7054	0.5851
0.56	35.8406	6200	0.7157	0.5858
0.5431	36.9971	6400	0.7262	0.5788
0.5281	38.1507	6600	0.6931	0.5598
0.5091	39.3072	6800	0.7102	0.5660
0.5031	40.4638	7000	0.6894	0.5466
0.4867	41.6203	7200	0.6858	0.5441
0.4689	42.7768	7400	0.7051	0.5403
0.4653	43.9333	7600	0.6840	0.5335
0.4441	45.0870	7800	0.6987	0.5322
0.435	46.2435	8000	0.7294	0.5315
0.4302	47.4	8200	0.6831	0.5182
0.4173	48.5565	8400	0.7080	0.5217
0.4086	49.7130	8600	0.6974	0.5101
0.3995	50.8696	8800	0.6842	0.5059
0.3866	52.0232	9000	0.7347	0.5186
0.3779	53.1797	9200	0.7141	0.5008
0.3691	54.3362	9400	0.7005	0.4998
0.3609	55.4928	9600	0.7299	0.4964
0.3584	56.6493	9800	0.6965	0.4966
0.3461	57.8058	10000	0.7217	0.4898
0.3403	58.9623	10200	0.7178	0.4850
0.3342	60.1159	10400	0.7019	0.4832
0.3215	61.2725	10600	0.7528	0.4834
0.3182	62.4290	10800	0.7112	0.4794
0.3123	63.5855	11000	0.7456	0.4780
0.3065	64.7420	11200	0.7509	0.4729
0.302	65.8986	11400	0.7293	0.4743
0.2942	67.0522	11600	0.7418	0.4734
0.2872	68.2087	11800	0.7607	0.4643
0.2844	69.3652	12000	0.7360	0.4679
0.2775	70.5217	12200	0.7594	0.4639
0.2736	71.6783	12400	0.7489	0.4667
0.2633	72.8348	12600	0.7576	0.4670
0.2627	73.9913	12800	0.7881	0.4597
0.2592	75.1449	13000	0.7566	0.4573
0.2557	76.3014	13200	0.7827	0.4629
0.246	77.4580	13400	0.7816	0.4586
0.2455	78.6145	13600	0.7918	0.4574
0.238	79.7710	13800	0.7928	0.4519
0.2376	80.9275	14000	0.7769	0.4508
0.2319	82.0812	14200	0.7877	0.4519
0.2268	83.2377	14400	0.7943	0.4537
0.2297	84.3942	14600	0.7913	0.4500
0.2207	85.5507	14800	0.8011	0.4481
0.219	86.7072	15000	0.7940	0.4485
0.2159	87.8638	15200	0.8179	0.4470
0.2126	89.0174	15400	0.8171	0.4449
0.2097	90.1739	15600	0.8208	0.4456
0.2074	91.3304	15800	0.8218	0.4446
0.2062	92.4870	16000	0.8242	0.4439
0.206	93.6435	16200	0.8340	0.4432
0.2007	94.8	16400	0.8240	0.4420
0.1994	95.9565	16600	0.8306	0.4419
0.197	97.1101	16800	0.8371	0.4431
0.1955	98.2667	17000	0.8345	0.4421
0.1979	99.4232	17200	0.8382	0.4421

Framework versions

• Transformers 4.49.0
• Pytorch 2.6.0+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0