README.md

metadata

license: apache-2.0
language: ar
datasets:
  - AymanMansour/SDN-Dialect-Dataset
  - arbml/sudanese_dialect_speech
  - arabic_speech_corpus
tags:
  - generated_from_trainer
metrics:
  - wer
model-index:
  - name: openai/whisper-small
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: SDN-Dialect-Dataset
          type: AymanMansour/SDN-Dialect-Dataset
          args: 'config: hi, split: test'
        metrics:
          - name: Wer
            type: wer
            value: 56.3216

openai/whisper-small

This model is a fine-tuned version of openai/whisper-small on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.5091
• Wer: 56.3216

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: 1e-05
• train_batch_size: 64
• 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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0157	13.0	1000	1.1631	65.9101
0.0025	26.0	2000	1.3416	58.5066
0.0009	39.01	3000	1.4238	56.6398
0.0004	52.01	4000	1.4800	56.3004
0.0002	65.01	5000	1.5091	56.3216

Framework versions

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