wav2vec2-xlsr-fi-lm-1B

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common voice train/dev/other datasets. It achieves the following results on the evaluation set without language model:

• Loss: 0.1853
• Wer: 0.2205

With language model:

• Wer: 0.1026

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: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.8158	0.67	400	0.4835	0.6310
0.5679	1.33	800	0.4806	0.5538
0.6055	2.0	1200	0.3888	0.5083
0.5353	2.67	1600	0.3258	0.4365
0.4883	3.33	2000	0.3313	0.4204
0.4513	4.0	2400	0.2924	0.3904
0.3753	4.67	2800	0.2593	0.3608
0.3478	5.33	3200	0.2832	0.3551
0.3796	6.0	3600	0.2495	0.3402
0.2556	6.67	4000	0.2342	0.3106
0.229	7.33	4400	0.2181	0.2812
0.205	8.0	4800	0.2041	0.2523
0.1654	8.67	5200	0.2015	0.2416
0.152	9.33	5600	0.1942	0.2294
0.1569	10.0	6000	0.1853	0.2205

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.11.0