whisper-sm-el-intlv-xl

This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 (el) and the google/fleurs (el_gr) datasets. It achieves the following results on the evaluation set:

• Loss: 0.4725
• Wer: 19.4837

Model description

The model was trained over 10000 steps on translation from Greek to English.

Intended uses & limitations

This model was part of the Whisper Finetuning Event (Dec 2022) and was used primarily to compare relative improvements between transcription and translation tasks.

Training and evaluation data

The training datasets combined examples from both train and evaluation splits and use the train split of the mozilla-foundation/common_voice_11_0 (el) dataset for evaluation and selection of the best checkpoint.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8.5e-06
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0545	2.49	1000	0.2891	22.4926
0.0093	4.98	2000	0.3927	20.1337
0.0018	7.46	3000	0.4031	20.1616
0.001	9.95	4000	0.4209	19.6880
0.0008	12.44	5000	0.4498	20.0966
0.0005	14.93	6000	0.4725	19.4837
0.0002	17.41	7000	0.4917	19.5951
0.0001	19.9	8000	0.5050	19.6230
0.0001	22.39	9000	0.5146	19.5672
0.0001	24.88	10000	0.5186	19.4837

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0
• Datasets 2.7.1.dev0
• Tokenizers 0.12.1