bangla_asr

This model is a fine-tuned version of Harveenchadha/vakyansh-wav2vec2-bengali-bnm-200 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 157.8652
• Wer: 0.4507

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: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 60
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2601.5363	7.46	500	259.6630	0.6863
417.7386	14.93	1000	156.6117	0.5275
262.9455	22.39	1500	155.0886	0.5006
178.7715	29.85	2000	155.1077	0.4840
132.448	37.31	2500	163.8623	0.4770
116.3943	44.78	3000	161.5531	0.4609
87.1653	52.24	3500	165.6857	0.4597
80.5606	59.7	4000	157.8652	0.4507

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.13.3
• Tokenizers 0.10.3