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

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: 0.9269
Wer: 0.5262

And on the test set:

Wer: 0.6225

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 minutes 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 minutes 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: 7e-05
train_batch_size: 16
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: 80
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
9.2929	6.25	50	3.0514	1.0
3.315	12.5	100	3.2255	1.0
3.1506	18.75	150	2.9924	1.0
2.9773	25.0	200	2.2199	1.0
2.1616	31.25	250	1.1423	0.8603
1.6887	37.5	300	0.9730	0.7020
1.1178	43.75	350	0.8971	0.6323
0.9512	50.0	400	0.9040	0.5960
0.7696	56.25	450	0.9232	0.5713
0.7348	62.5	500	0.9203	0.5412
0.9312	68.75	550	0.9673	0.5376
0.6519	75.0	600	0.9269	0.5262

Framework versions

Transformers 4.28.1
Pytorch 2.0.0+cu117
Datasets 2.11.0
Tokenizers 0.13.3