whisper-small-ar-v2

This model is for Arabic automatic speech recognition (ASR). It is a fine-tuned version of openai/whisper-small on the Arabic portion of the mozilla-foundation/common_voice_16_1 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4007
• Wer: 47.7264

Model description

Whisper model fine-tuned on Arabic data, following the official tutorial.

Intended uses & limitations

It is recommended to fine-tune and evaluate on your data before using it.

Training and evaluation data

Training Data: CommonVoice (v16.1) Arabic train + validation splits
Validation Data: CommonVoice (v16.1) Arabic test split

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 32
• eval_batch_size: 32
• 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: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2742	0.82	1000	0.3790	275.2463
0.1625	1.65	2000	0.3353	228.5252
0.1002	2.47	3000	0.3311	238.8858
0.0751	3.3	4000	0.3354	158.1532
0.0601	4.12	5000	0.3576	48.9285
0.0612	4.95	6000	0.3575	47.8937
0.0383	5.77	7000	0.3819	46.9085
0.0234	6.6	8000	0.4007	47.7264

Framework versions

• Transformers 4.38.1
• Pytorch 2.1.0+cu118
• Datasets 2.17.1
• Tokenizers 0.15.2