Model Card for RawandLaouini/ArabicEchoV2

Model Details

Model Description

This is a fine-tuned version of the openai/whisper-medium model, adapted for Arabic Automatic Speech Recognition (ASR) using LoRA (Low-Rank Adaptation). The model was trained on the custom Whisper_Arabic_Merged_v6 dataset, containing 1183 audio samples, to improve transcription accuracy for Arabic speech.

Developed by: Rawand Laouini
Finetuned from model: openai/whisper-medium
Model type: Transformer-based ASR model with LoRA
Language(s): Arabic
License: MIT (or specify your preferred license)
Shared by: Rawand Laouini

Uses

Direct Use

This model can be used for transcribing Arabic speech to text, ideal for applications like voice assistants, subtitle generation, or educational tools tailored to Arabic speakers.

Out-of-Scope Use

The model should not be used for real-time transcription without optimization, nor for languages other than Arabic without retraining.

Bias, Risks, and Limitations

Trained on the Whisper_Arabic_Merged_v6 dataset, the model may reflect biases or limitations in dialectal coverage or audio quality. Performance may vary with different Arabic dialects or noisy conditions. Users should validate outputs for critical use.

Recommendations

Test the model on your specific use case and consider expanding the dataset for better dialectal or noise robustness.

How to Get Started with the Model

Use the following code to load and use the model:

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel

processor = WhisperProcessor.from_pretrained("openai/whisper-medium")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-medium")
model = PeftModel.from_pretrained(model, "RawandLaouini/ArabicEchoV2")

model.eval()
input_features = processor(audio, return_tensors="pt").input_features
predicted_ids = model.generate(input_features)
transcription = processor.tokenizer.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Training Details

Training Data

Dataset: RawandLaouini/Whisper_Arabic_Merged_v6 (1183 samples)
Training split: 946 examples
Validation split: Manual evaluation on 50 examples

Training Procedure

Preprocessing

Audio was processed to match Whisper's requirements, with input features extracted using the Whisper processor.

Training Hyperparameters

Batch size: 1 (per device)
Gradient accumulation steps: 1
Learning rate: 1e-4
Warmup steps: 100
Max steps: 300
Optimizer: AdamW
Mixed precision: FP16

Speeds, Sizes, Times

Training time: ~2.43 minutes for 300 steps
Model size: Lightweight (LoRA adapters)

Evaluation

Testing Data

Manual evaluation on 50 examples from the validation split.

Metrics

Word Error Rate (WER): 0.2969
Character Error Rate (CER): 0.0700

Results

The model achieves a WER of 29.69% and CER of 7.00% on the manual evaluation set, indicating good transcription accuracy for Arabic speech.

Environmental Impact

Hardware Type: NVIDIA GPU (14.74 GiB)
Hours used: ~0.04 hours (2.43 minutes)
Cloud Provider: Local/Colab (unspecified)
Compute Region: Unspecified
Carbon Emitted: Minimal (estimated < 0.01 kg CO2e using Lacoste et al., 2019)

Citation

BibTeX:

@misc{laouini2025arabicechov2,
  author = {Rawand Laouini},
  title = {ArabicEchoV2: Fine-tuned Whisper-medium for Arabic ASR with LoRA},
  year = {2025},
  howpublished = {\url{https://huggingface.co/RawandLaouini/ArabicEchoV2}}
}

APA:

Laouini, R. (2025). ArabicEchoV2: Fine-tuned Whisper-medium for Arabic ASR with LoRA. Retrieved from https://huggingface.co/RawandLaouini/ArabicEchoV2

Model Card Authors

Rawand Laouini

Model Card Contact

Email: [awini.rawand21@gmail.com]