README.md

---
language:
- ar
license: apache-2.0
task_categories:
- automatic-speech-recognition
tags:
- whisper
- arabic
- speech
- asr
- multidialect
pretty_name: Arabic Whisper Multi-Dialect (Processed) - Small
size_categories:
- 10K<n<100K
---

# Arabic Whisper Multi-Dialect - Processed (Small)

## Dataset Description

This is a **preprocessed** version of the Arabic multi-dialect speech dataset, ready for fine-tuning OpenAI's Whisper models. The dataset contains audio features extracted and formatted specifically for Whisper training.

- **Size**: 40% subset of the full `arabic-whisper-multidialect` dataset
- **Total Examples**: 43,091 samples
- **Format**: Pre-computed Whisper input features (mel spectrograms) and tokenized labels
- **Purpose**: Direct use with Whisper training pipelines without additional preprocessing

### Key Features

✅ **Pre-processed** - Audio already converted to Whisper input features  
✅ **Multi-dialect** - Covers various Arabic dialects  
✅ **Training-ready** - No additional feature extraction needed  
✅ **Memory-efficient** - Optimized for faster loading during training  

## Dataset Structure

### Data Splits

| Split | Examples | Size (GB) |
|-------|----------|-----------|
| Train | 37,835 | 33.8 |
| Validation | 2,628 | 2.35 |
| Test | 2,628 | 2.35 |
| **Total** | **43,091** | **38.5** |

### Data Fields

- **`input_features`**: `Sequence[Sequence[float32]]`
  - Shape: `(80, 3000)` - 80 mel-frequency bins × 3000 time steps
  - Pre-computed log-mel spectrogram features for Whisper
  
- **`labels`**: `Sequence[int64]`
  - Tokenized Arabic transcription using Whisper's tokenizer
  - Special tokens: `-100` for padding (ignored in loss computation)

## Usage

### Quick Start

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("MadLook/arabic-whisper-multidialect-processed-small")

print(dataset)
# DatasetDict({
#     train: Dataset with 37,835 examples
#     validation: Dataset with 2,628 examples
#     test: Dataset with 2,628 examples
# })
```


### Loading a Single Split

```python
# Load only training data
train_data = load_dataset("MadLook/arabic-whisper-multidialect-processed-small", split="train")

# Load with streaming for large datasets
train_stream = load_dataset("MadLook/arabic-whisper-multidialect-processed-small", split="train", streaming=True)
```

## Preprocessing Details

This dataset was created from the original Arabic multi-dialect dataset using the following preprocessing pipeline:

1. **Audio Loading**: Resampled to 16kHz mono audio
2. **Feature Extraction**: Converted to 80-channel log-mel spectrograms using Whisper's feature extractor
3. **Tokenization**: Arabic text transcriptions tokenized with Whisper's multilingual tokenizer
4. **Normalization**: Applied Whisper's standard audio normalization
5. **Subset Selection**: Selected 40% of the original dataset

### Processing Code

```python
from transformers import WhisperProcessor

processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="ar", task="transcribe")

def prepare_dataset(batch):
    # Load and resample audio
    audio = batch["audio"]
    
    # Compute input features
    batch["input_features"] = processor.feature_extractor(
        audio["array"], 
        sampling_rate=audio["sampling_rate"]
    ).input_features[0]
    
    # Tokenize transcription
    batch["labels"] = processor.tokenizer(batch["transcription"]).input_ids
    
    return batch
```

## Source Dataset

This is a processed subset of the Arabic multi-dialect speech recognition dataset, which includes recordings from various Arabic-speaking regions and dialects.

## Intended Use

### Primary Use Cases

- Fine-tuning Whisper models for Arabic speech recognition
- Research on multi-dialect Arabic ASR
- Benchmarking Arabic speech recognition systems
- Transfer learning for low-resource Arabic dialects

### Out-of-Scope Use

- This dataset is **already preprocessed** - do not apply feature extraction again
- Not suitable for training non-Whisper architectures without re-processing

## Limitations

- Only 40% of the full dataset (subset for faster experimentation)
- Pre-computed features are specific to Whisper architecture
- Fixed audio length (30 seconds max due to Whisper constraints)


## Citation

If you use this dataset, please cite both this preprocessed version and the original source dataset:

```bibtex
@dataset{arabic_whisper_multidialect_processed,
  title={Arabic Whisper Multi-Dialect - Processed (Small)},
  author={MadLook},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/MadLook/arabic-whisper-multidialect-processed-small}
}
```

## License

Apache 2.0

## Contact

For questions or issues with this dataset, please open an issue on the dataset repository.

---

**Dataset Version**: 1.0  
**Last Updated**: 2025  
**Preprocessor**: Whisper Feature Extractor (openai/whisper-small)  
**Compatible Models**: openai/whisper-tiny, openai/whisper-base, openai/whisper-small, openai/whisper-medium, openai/whisper-large