whisper-md-el-intlv-xs

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

• Loss: 0.4168
• Wer: 11.3670

Model description

This model is trained over the two interleaved datasets in the Greek language. Testing used only the common_voice_11_0 (el) test split.

Intended uses & limitations

The model was trained for transcription in Greek

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-06
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• 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
0.0251	2.49	1000	0.2216	12.5836
0.0051	4.98	2000	0.2874	12.2957
0.0015	7.46	3000	0.3281	11.9056
0.0017	9.95	4000	0.3178	12.5929
0.0008	12.44	5000	0.3449	11.9799
0.0001	14.93	6000	0.3638	11.7106
0.0001	17.41	7000	0.3910	11.4970
0.0	19.9	8000	0.4042	11.3949
0.0	22.39	9000	0.4129	11.4134
0.0	24.88	10000	0.4168	11.3670

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2