gha-whisper-small-twi-v3

This model is a fine-tuned version of teckedd/whisper-small-waxal-round2-specaug-v1 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.2959
• Wer: 0.2807
• Cer: 0.0958

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: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• training_steps: 1500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.4513	0.4747	150	0.4382	0.3044	0.1058
0.308	0.9494	300	0.3023	0.2897	0.0975
0.2678	1.4241	450	0.2996	0.2969	0.1042
0.2636	1.8987	600	0.2959	0.2807	0.0958
0.2081	2.3734	750	0.2998	0.2896	0.0994
0.2006	2.8481	900	0.2989	0.2932	0.1012
0.1614	3.3228	1050	0.3043	0.3010	0.1050
0.169	3.7975	1200	0.3035	0.2906	0.1009
0.152	4.2722	1350	0.3090	0.3004	0.1065
0.1512	4.7468	1500	0.3084	0.2984	0.1034

Framework versions

• Transformers 4.46.3
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3