LegolasTheElf/Wav2Vec2_xls_r_lm_300m_hi

Wav2Vec2_xls_r_300m_hi_final

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the 'Openslr Multilingual and code-switching ASR challenge' dataset and 'mozilla-foundation/common_voice_7_0' dataset. It achieves the following results on the evaluation set:

• Loss: 0.3035
• Wer: 0.3137
• Cer: 0.0972

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: 16
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 8
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.9821	0.64	400	0.5059	0.4783	0.1573
0.6861	1.28	800	0.4201	0.4247	0.1356
0.585	1.92	1200	0.3797	0.3811	0.1210
0.5193	2.56	1600	0.3577	0.3652	0.1152
0.4583	3.21	2000	0.3422	0.3519	0.1111
0.4282	3.85	2400	0.3261	0.3450	0.1071
0.3951	4.49	2800	0.3201	0.3325	0.1048
0.3619	5.13	3200	0.3167	0.3296	0.1030
0.345	5.77	3600	0.3157	0.3210	0.1013
0.338	6.41	4000	0.3051	0.3143	0.0982
0.3155	7.05	4400	0.3059	0.3154	0.0986
0.3057	7.69	4800	0.3035	0.3137	0.0972

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0