amenIKh/Tunisian_Checkpoint12

Model Card for Model ID

Model Details

Model Description

Model Card for Model ID

Finetuning Whisper on Tunisian custom dataset

Model Details

Model Description

This model is a fine-tuned version of openai/whisper-small on the tunisian_custom dataset =more than 4h(/doumawl4+/doumaw02+Data3+dataset1+dataset2). It achieves the following results on the evaluation set:

• Train Loss: 0.0109
• Evaluation Loss: 1.1608097553253174
• Wer: 54.52300785634119
• Cer: 25.538666370797735 -max_audio_length=15 for the preprocessing i used padding+VAD filter
• Developed by: [Ameni Khabthani]
• Funded by [optional]: [More Information Needed]
• Shared by [optional]: [More Information Needed]
• Model type: [ASR system]
• Language(s) (NLP): [More Information Needed]
• License: [More Information Needed]
• Finetuned from model [optional]: [whisper small]

Model Sources [optional]

• Repository: [More Information Needed]
• Paper [optional]: [More Information Needed]
• Demo [optional]: [More Information Needed]

Training Hyperparameters

per_device_train_batch_size=4
gradient_accumulation_steps=8
learning_rate= 5e-5

warmup_steps=100
max_steps=4000
gradient_checkpointing=True
fp16=True
save_steps=500
eval_steps=500
per_device_eval_batch_size=8
predict_with_generate=True
generation_max_length=249
logging_steps=50
weight_decay=0.001 dropout=0.1 optim="adamw_bnb_8bit"
seed=42 save_total_limit=5 save_steps=500,
eval_steps=500,