RobbieJimersonJr/whisper-medium-Bambara-field

Whisper Medium Bambara Fieldwork

This model is a fine-tuned version of openai/whisper-medium on the audiofolder dataset. It achieves the following results on the evaluation set:

• Loss: 4.0725
• Wer: 157.6004

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: 1e-05
• train_batch_size: 16
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 13532
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6342	1.03	1000	2.5810	159.5293
0.2873	3.03	2000	2.9513	159.1140
0.1461	5.02	3000	3.6833	158.3941
0.049	7.02	4000	4.0725	157.6004
0.0218	9.01	5000	4.2531	158.3664
0.0071	11.0	6000	4.5944	157.9972
0.0057	12.04	7000	4.6659	161.4952
0.0061	14.03	8000	4.9162	161.0614
0.0042	16.03	9000	5.0205	158.9848
0.0007	18.02	10000	5.1463	159.1970
0.0015	20.01	11000	5.2150	159.0401
0.0001	22.01	12000	5.3008	159.5478
0.0001	24.0	13000	5.3800	159.0309

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.1+cu117
• Datasets 2.8.0
• Tokenizers 0.13.2