wav2vec2-mms-1b-CV17.0

This model is a fine-tuned version of facebook/mms-1b-all on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.6919
• Wer: 0.6538
• Cer: 0.2510
• Bleu: 0.1420

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: 0.001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.15
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Bleu
6.2938	3.0769	200	3.8350	0.9981	0.9092	0.0
2.0522	6.1538	400	0.7219	0.6997	0.2730	0.1116
0.7043	9.2308	600	0.7137	0.7419	0.2682	0.0933
0.6497	12.3077	800	0.6962	0.6664	0.2667	0.1318
0.614	15.3846	1000	0.6680	0.6586	0.2596	0.1356
0.5794	18.4615	1200	0.6798	0.6722	0.2599	0.1254
0.5439	21.5385	1400	0.6724	0.6665	0.2541	0.1287
0.5146	24.6154	1600	0.6906	0.6704	0.2513	0.1327
0.489	27.6923	1800	0.6886	0.6599	0.2509	0.1390
0.4668	30.7692	2000	0.6919	0.6538	0.2510	0.1420

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.0+cu121
• Datasets 3.0.0
• Tokenizers 0.19.1