pranetk/whisper-largev3-paraspeak-v2

Whisper Large V3 Paraspeak V2

This model is a fine-tuned version of openai/whisper-large-v3 on the Paraspeak Dataset 3.0 dataset. It achieves the following results on the evaluation set:

• Loss: 1.0662
• Wer: 62.1212

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: 3e-05
• 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
• training_steps: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1361	2.9412	50	0.8743	77.2727
0.0603	5.8824	100	1.0115	65.1515
0.0452	8.8235	150	1.0837	71.2121
0.0042	11.7647	200	1.0400	78.7879
0.023	14.7059	250	1.0296	71.2121
0.0023	17.6471	300	0.9761	69.6970
0.0005	20.5882	350	1.0758	71.2121
0.0098	23.5294	400	1.1036	71.2121
0.0006	26.4706	450	1.0662	65.1515
0.0001	29.4118	500	1.0563	62.1212
0.0	32.3529	550	1.0521	62.1212
0.0	35.2941	600	1.0541	62.1212
0.0	38.2353	650	1.0563	62.1212
0.0	41.1765	700	1.0587	62.1212
0.0	44.1176	750	1.0609	62.1212
0.0	47.0588	800	1.0628	62.1212
0.0	50.0	850	1.0641	62.1212
0.0	52.9412	900	1.0653	62.1212
0.0	55.8824	950	1.0659	62.1212
0.0	58.8235	1000	1.0662	62.1212

Framework versions

• Transformers 4.42.3
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1