wav2vec2-large-xls-r-1b-frisian-cv-8-10h

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice_8_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1207
• Wer: 0.0961

And on the test set:

• Wer: 0.0883

Model description

This model has been developed for my Master's thesis in "Voice Technology" at Rijksuniversiteit Groningen - Campus Fryslân. It corresponds to experiment 3 where I use as training set 10 hours of Frisian speech randomly selected from all validated data except the test and evaluation sets.

Intended uses & limitations

The intended use is for recognizing Frisian speech.

Limitations include no LM rescoring and using version 8.0 of Common Voice instead of 13.0.

Training and evaluation data

The evaluation split used is the one available in the Common Voice 8.0 Frisian subset. The train split is 10 hours of Frisian randomly selected from validated data except for the recordings from test and evaluation splits.

Training procedure

The script used for training this model can be found in this GitHub repository: link.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.6342	1.32	300	2.9760	1.0
2.2716	2.63	600	0.6877	0.6024
1.1303	3.95	900	0.3522	0.3450
0.9038	5.26	1200	0.2714	0.2603
0.846	6.58	1500	0.2143	0.2036
0.8044	7.89	1800	0.1829	0.1788
0.7069	9.21	2100	0.1751	0.1667
0.6995	10.53	2400	0.1741	0.1727
0.7115	11.84	2700	0.1591	0.1486
0.677	13.16	3000	0.1636	0.1459
0.6032	14.47	3300	0.1535	0.1439
0.6218	15.79	3600	0.1427	0.1406
0.6519	17.11	3900	0.1498	0.1488
0.5739	18.42	4200	0.1438	0.1319
0.567	19.74	4500	0.1379	0.1322
0.4982	21.05	4800	0.1315	0.1237
0.5825	22.37	5100	0.1349	0.1252
0.5085	23.68	5400	0.1297	0.1233
0.4946	25.0	5700	0.1343	0.1127
0.5677	26.32	6000	0.1323	0.1228
0.4858	27.63	6300	0.1292	0.1098
0.4709	28.95	6600	0.1267	0.1204
0.3241	30.26	6900	0.1315	0.1274
0.2796	31.58	7200	0.1315	0.1202
0.3171	32.89	7500	0.1315	0.1200
0.2591	34.21	7800	0.1322	0.1106
0.2716	35.53	8100	0.1233	0.1030
0.2446	36.84	8400	0.1273	0.1087
0.2377	38.16	8700	0.1243	0.1101
0.2183	39.47	9000	0.1230	0.1116
0.2059	40.79	9300	0.1240	0.1001
0.1916	42.11	9600	0.1223	0.1003
0.196	43.42	9900	0.1246	0.0965
0.1969	44.74	10200	0.1222	0.1038
0.1951	46.05	10500	0.1208	0.1003
0.1809	47.37	10800	0.1213	0.1003
0.1793	48.68	11100	0.1202	0.0959
0.1837	50.0	11400	0.1207	0.0961

Framework versions

• Transformers 4.28.1
• Pytorch 2.0.0+cu117
• Datasets 2.11.0
• Tokenizers 0.13.3