Kalemat-Tech Arabic Speech Recognition Model (STT) - Mohamed Salama

نموذج كلماتك للتعرف على الأصوات العربية الفصحى و تحويلها إلى نصوص

KalemaTech-Arabic-STT-ASR-based-on-Whisper-Small

This model is a fine-tuned version of openai/whisper-small on Common_Voice_Arabic_12.0_Augmented. It achieves the following results on the evaluation set:

• Loss: 0.5362
• Wer: 58.5848

Example of usage:

from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq

processor = AutoProcessor.from_pretrained("Salama1429/KalemaTech-Arabic-STT-ASR-based-on-Whisper-Small")

model = AutoModelForSpeechSeq2Seq.from_pretrained("Salama1429/KalemaTech-Arabic-STT-ASR-based-on-Whisper-Small")

Intended uses & limitations

Automatic Speech Recognition

Training and evaluation data

Common_Voice_Arabic_12.0 and I made some augmentations to it as follows:
- 25% of the data TimeMasking
- 25% of the data SpecAugmentation
- 25% of the data WavAugmentation (AddGaussianNoise)
- The final dataset is the original common voice plus the augmented files

Training procedure

Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 64
- 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
- num_epochs: 25
- mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.1+cu117
• Datasets 2.8.0
• Tokenizers 0.13.2