farsipal/whisper-sm-el-intlv-xs

Whisper small (Greek) Trained on Interleaved Datasets

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

• Loss: 0.4741
• Wer: 20.0687

Model description

The model was developed during the Whisper Fine-Tuning Event in December 2022. More details on the model can be found in the original paper

Intended uses & limitations

The model is fine-tuned for transcription in the Greek language.

Training and evaluation data

This model was trained by interleaving the training and evaluation splits from two different datasets:

• mozilla-foundation/common_voice_11_0 (el)
• google/fleurs (el_gr)

Training procedure

The python script used is a modified version of the script provided by Hugging Face and can be found here

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0186	4.98	1000	0.3619	21.0067
0.0012	9.95	2000	0.4347	20.3009
0.0005	14.93	3000	0.4741	20.0687
0.0003	19.9	4000	0.4974	20.1152
0.0003	24.88	5000	0.5066	20.2266

Framework versions

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