qmeeus/whisper-small-multilingual-spoken-ner-end2end-v2

WhisperForSpokenNER-end2end

This model is a fine-tuned version of openai/whisper-small on the facebook/voxpopuli de+es+fr+nl dataset. It achieves the following results on the evaluation set:

• Loss: 0.2755
• Combined Wer: 0.1491
• F1 Score: 0.7163
• Label F1: 0.8200
• Wer: 0.0858

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: 64
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Combined Wer	F1 Score	Label F1	Wer
0.3252	0.1	500	0.3396	0.1918	0.6148	0.7578	0.1193
0.2729	0.2	1000	0.3158	0.1730	0.6449	0.7907	0.1058
0.2369	0.3	1500	0.2971	0.1736	0.6917	0.8083	0.1067
0.1967	0.4	2000	0.2823	0.1634	0.6915	0.8095	0.0999
0.1623	0.5	2500	0.2804	0.1693	0.7088	0.8249	0.1052
0.1146	1.02	3000	0.2820	0.1593	0.7012	0.8106	0.0951
0.0938	1.12	3500	0.2792	0.1500	0.7205	0.8238	0.0875
0.1001	1.22	4000	0.2750	0.1549	0.7072	0.8061	0.0928
0.0848	1.32	4500	0.2741	0.1471	0.7243	0.8318	0.0860
0.0649	1.42	5000	0.2745	0.1468	0.7304	0.8350	0.0858

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.1.0
• Datasets 2.14.6
• Tokenizers 0.14.1