wav2vec2-large-xlsr-common1000asli-demo-colab-dd

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 1.0671
• Wer: 0.5268

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: 0.0003
• train_batch_size: 128
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 256
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
10.4469	10.53	400	2.9884	0.9788
1.6358	21.05	800	0.5968	0.6578
0.3404	31.58	1200	0.5348	0.5757
0.2218	42.11	1600	0.5945	0.5659
0.1676	52.63	2000	0.6318	0.5753
0.1442	63.16	2400	0.6395	0.5559
0.1237	73.68	2800	0.6981	0.5522
0.1132	84.21	3200	0.7174	0.5428
0.1056	94.74	3600	0.7120	0.5531
0.0942	105.26	4000	0.7865	0.5529
0.0967	115.79	4400	0.7796	0.5546
0.0854	126.32	4800	0.7392	0.5507
0.0816	136.84	5200	0.8173	0.5575
0.0748	147.37	5600	0.8164	0.5550
0.0691	157.89	6000	0.8061	0.5444
0.0654	168.42	6400	0.8098	0.5524
0.0627	178.95	6800	0.8527	0.5655
0.0653	189.47	7200	0.8210	0.5412
0.0639	200.0	7600	0.8619	0.5602
0.0614	210.53	8000	0.8453	0.5606
0.0647	221.05	8400	0.8248	0.5564
0.0611	231.58	8800	0.8323	0.5637
0.0606	242.11	9200	0.8754	0.5654
0.0587	252.63	9600	0.8684	0.5528
0.0524	263.16	10000	0.8798	0.5556
0.0499	273.68	10400	0.8593	0.5553
0.0466	284.21	10800	0.9079	0.5520
0.0505	294.74	11200	0.8680	0.5607
0.0517	305.26	11600	0.8709	0.5557
0.0522	315.79	12000	0.8687	0.5570
0.0453	326.32	12400	0.8585	0.5614
0.047	336.84	12800	0.9249	0.5581
0.0431	347.37	13200	0.8934	0.5543
0.0454	357.89	13600	0.8837	0.5583
0.0472	368.42	14000	0.9070	0.5565
0.0431	378.95	14400	0.9202	0.5526
0.0404	389.47	14800	0.9234	0.5543
0.0386	400.0	15200	0.9056	0.5549
0.0372	410.53	15600	0.9901	0.5493
0.0376	421.05	16000	0.9109	0.5460
0.0365	431.58	16400	0.9313	0.5487
0.0347	442.11	16800	0.9027	0.5496
0.0361	452.63	17200	0.9614	0.5457
0.0323	463.16	17600	0.9782	0.5558
0.0325	473.68	18000	0.9549	0.5481
0.032	484.21	18400	0.9781	0.5431
0.0289	494.74	18800	0.9840	0.5463
0.0292	505.26	19200	0.9397	0.5357
0.0276	515.79	19600	0.9228	0.5467
0.0283	526.32	20000	0.9683	0.5394
0.0281	536.84	20400	0.9783	0.5479
0.026	547.37	20800	0.9663	0.5472
0.0288	557.89	21200	0.9424	0.5426
0.0275	568.42	21600	0.9788	0.5435
0.0264	578.95	22000	0.9703	0.5473
0.0259	589.47	22400	0.9994	0.5446
0.0243	600.0	22800	0.9637	0.5590
0.0251	610.53	23200	0.9577	0.5457
0.0222	621.05	23600	0.9780	0.5419
0.0227	631.58	24000	0.9582	0.5417
0.0222	642.11	24400	0.9847	0.5432
0.0214	652.63	24800	1.0171	0.5449
0.022	663.16	25200	0.9819	0.5430
0.0202	673.68	25600	0.9737	0.5413
0.0187	684.21	26000	0.9977	0.5440
0.0213	694.74	26400	0.9919	0.5464
0.0197	705.26	26800	0.9769	0.5357
0.0183	715.79	27200	0.9964	0.5377
0.0187	726.32	27600	0.9973	0.5341
0.0191	736.84	28000	0.9970	0.5399
0.0183	747.37	28400	1.0179	0.5371
0.0176	757.89	28800	1.0020	0.5440
0.018	768.42	29200	0.9992	0.5394
0.0157	778.95	29600	1.0502	0.5397
0.0165	789.47	30000	1.0463	0.5397
0.0147	800.0	30400	1.0363	0.5430
0.0153	810.53	30800	0.9890	0.5407
0.0145	821.05	31200	1.0139	0.5369
0.0143	831.58	31600	1.0260	0.5346
0.0141	842.11	32000	1.0277	0.5361
0.0139	852.63	32400	1.0639	0.5335
0.0132	863.16	32800	1.0661	0.5314
0.013	873.68	33200	1.0537	0.5335
0.0126	884.21	33600	1.0433	0.5347
0.0121	894.74	34000	1.0275	0.5326
0.0128	905.26	34400	1.0405	0.5327
0.0112	915.79	34800	1.0626	0.5296
0.0115	926.32	35200	1.0583	0.5284
0.0109	936.84	35600	1.0494	0.5287
0.0113	947.37	36000	1.0655	0.5294
0.0104	957.89	36400	1.0723	0.5269
0.0108	968.42	36800	1.0680	0.5267
0.0104	978.95	37200	1.0707	0.5261
0.0108	989.47	37600	1.0649	0.5268
0.0103	1000.0	38000	1.0671	0.5268

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu102
• Datasets 1.13.3
• Tokenizers 0.10.3