wav2vec2-xlsr-53-300m-mls-german-ft

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the MULTILINGUAL_LIBRISPEECH - GERMAN 10h dataset. It achieves the following results on the evaluation set:

• Loss: 0.2219
• Wer: 0.1288

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: 200.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9888	7.25	500	2.9192	1.0
2.9313	14.49	1000	2.8698	1.0
1.068	21.74	1500	0.2647	0.2565
0.8151	28.99	2000	0.2067	0.1719
0.764	36.23	2500	0.1975	0.1568
0.7332	43.48	3000	0.1812	0.1463
0.5952	50.72	3500	0.1923	0.1428
0.6655	57.97	4000	0.1900	0.1404
0.574	65.22	4500	0.1822	0.1370
0.6211	72.46	5000	0.1937	0.1355
0.5883	79.71	5500	0.1872	0.1335
0.5666	86.96	6000	0.1874	0.1324
0.5526	94.2	6500	0.1998	0.1368
0.5671	101.45	7000	0.2054	0.1365
0.5514	108.7	7500	0.1987	0.1340
0.5382	115.94	8000	0.2104	0.1344
0.5819	123.19	8500	0.2125	0.1334
0.5277	130.43	9000	0.2063	0.1330
0.4626	137.68	9500	0.2105	0.1310
0.5842	144.93	10000	0.2087	0.1307
0.535	152.17	10500	0.2137	0.1309
0.5081	159.42	11000	0.2215	0.1302
0.6033	166.67	11500	0.2162	0.1302
0.5549	173.91	12000	0.2198	0.1286
0.5389	181.16	12500	0.2241	0.1293
0.4912	188.41	13000	0.2190	0.1290
0.4671	195.65	13500	0.2218	0.1290

Framework versions

• Transformers 4.13.0.dev0
• Pytorch 1.10.0
• Datasets 1.15.2.dev0
• Tokenizers 0.10.3