Upper Sorbian wav2vec2-xls-r-300m-hsb-cv8

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

• Loss: 0.9643
• Wer: 0.5037
• Cer: 0.1278

Evaluation

The model can be evaluated using the attached eval.py script:

python eval.py --model_id comodoro/wav2vec2-xls-r-300m-hsb-cv8 --dataset mozilla-foundation/common-voice_8_0 --split test --config hsb

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 16
• 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: 200
• num_epochs: 500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
4.3121	19.35	1200	3.2059	1.0	1.0
2.6525	38.71	2400	1.1324	0.9387	0.3204
1.3644	58.06	3600	0.8767	0.8099	0.2271
1.093	77.42	4800	0.8739	0.7603	0.2090
0.9546	96.77	6000	0.8454	0.6983	0.1882
0.8554	116.13	7200	0.8197	0.6484	0.1708
0.775	135.48	8400	0.8452	0.6345	0.1681
0.7167	154.84	9600	0.8551	0.6241	0.1631
0.6609	174.19	10800	0.8442	0.5821	0.1531
0.616	193.55	12000	0.8892	0.5864	0.1527
0.5815	212.9	13200	0.8839	0.5772	0.1503
0.55	232.26	14400	0.8905	0.5665	0.1436
0.5173	251.61	15600	0.8995	0.5471	0.1417
0.4969	270.97	16800	0.8633	0.5325	0.1334
0.4803	290.32	18000	0.9074	0.5253	0.1352
0.4596	309.68	19200	0.9159	0.5146	0.1294
0.4415	329.03	20400	0.9055	0.5189	0.1314
0.434	348.39	21600	0.9435	0.5208	0.1314
0.4199	367.74	22800	0.9199	0.5136	0.1290
0.4008	387.1	24000	0.9342	0.5174	0.1303
0.4051	406.45	25200	0.9436	0.5132	0.1292
0.3861	425.81	26400	0.9417	0.5084	0.1283
0.3738	445.16	27600	0.9573	0.5079	0.1299
0.3768	464.52	28800	0.9682	0.5062	0.1289
0.3647	483.87	30000	0.9643	0.5037	0.1278

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0