scream_medium_beta

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

• step: 24999
• validation_fleurs_loss: 1.4171
• train_loss: 0.5400
• validation_fleurs_wer: 8.8638
• validation_fleurs_cer: 3.8370
• validation_fleurs_exact_wer: 14.1278
• validation_fleurs_exact_cer: 5.1993
• validation_stortinget_loss: 0.3369
• validation_stortinget_wer: 14.2120
• validation_stortinget_cer: 10.2972
• validation_stortinget_exact_wer: 17.4640
• validation_stortinget_exact_cer: 10.8352
• validation_nrk_tv_loss: 0.8259
• validation_nrk_tv_wer: 39.9035
• validation_nrk_tv_cer: 31.1762
• validation_nrk_tv_exact_wer: 47.4289
• validation_nrk_tv_exact_cer: 32.3674

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: 2.5e-05
• lr_scheduler_type: linear
• per_device_train_batch_size: 16
• total_train_batch_size_per_node: 64
• total_train_batch_size: 1024
• total_optimization_steps: 25,000
• starting_optimization_step: None
• finishing_optimization_step: 25,000
• num_train_dataset_workers: 32
• num_hosts: 16
• total_num_training_examples: 25,600,000
• steps_per_epoch: 6271
• num_beams: None
• dropout: True
• bpe_dropout_probability: 0.1

Training results

step	validation_fleurs_loss	train_loss	validation_fleurs_wer	validation_fleurs_cer	validation_fleurs_exact_wer	validation_fleurs_exact_cer	validation_stortinget_loss	validation_stortinget_wer	validation_stortinget_cer	validation_stortinget_exact_wer	validation_stortinget_exact_cer	validation_nrk_tv_loss	validation_nrk_tv_wer	validation_nrk_tv_cer	validation_nrk_tv_exact_wer	validation_nrk_tv_exact_cer
0	3.6595	2.4764	17.4301	5.4794	21.6249	6.3977	1.3465	33.9515	19.1377	38.4072	20.3275	1.8386	66.2133	48.0904	75.6490	49.8313
5000	1.2828	0.6841	8.8638	3.8864	13.2318	4.9529	0.3311	14.5798	10.4664	17.7786	11.0119	0.8229	41.0824	31.7759	48.7519	33.0088
10000	1.1134	0.6019	8.3284	3.6990	13.1123	4.9287	0.3132	14.0485	10.1394	17.2099	10.6785	0.7856	39.0957	30.3740	46.7798	31.5896
15000	1.1605	0.5821	8.5068	3.7631	13.5603	5.0157	0.3181	13.7633	10.0236	16.9465	10.5585	0.7864	39.4419	30.9142	46.7507	32.1012
20000	1.0986	0.5395	8.7448	3.9456	14.4863	5.3733	0.3226	14.2469	10.3402	17.4640	10.8776	0.7884	39.8129	31.0325	47.3332	32.2280

Framework versions

• Transformers 4.31.0.dev0
• Datasets 2.13.0
• Tokenizers 0.13.3