---
language:
- ar
- en
pretty_name: Synthetic Arabic-English Code-Switched Speech for ASR
tags:
- audio
- automatic-speech-recognition
- code-switching
- egyptian-arabic
- synthetic-speech
- datasets
task_categories:
- automatic-speech-recognition
size_categories:
- 1K<n<10K
license: other
configs:
- config_name: synthetic
  data_files:
  - split: train
    path: data/synthetic/train-*.parquet
  - split: test
    path: data/synthetic/test-*.parquet
---

# Synthetic Arabic-English Code-Switched Speech for ASR

This dataset contains synthetic speech generated for Egyptian Arabic-English code-switched automatic speech recognition. It is published separately from the human review annotations so the human audio remains in its [upstream Hugging Face repository](https://huggingface.co/datasets/MohamedRashad/arabic-english-code-switching).

## Configurations

| Configuration | Train | Test | Publication status |
|---|---:|---:|---|
| `synthetic` | 8,655 | 962 | Contains 5,673 ArE-CSTD-derived texts; noncommercial/share-alike terms apply |

The 3,944 author-created CC BY 4.0 rows remain identifiable through the `license` and `source_type` fields. Keeping one Parquet configuration avoids storing the same lossless audio twice.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("abdo1819/arabic-english-code-switching-synthetic-asr", "synthetic")

generated_only = dataset.filter(lambda row: row["license"] == "cc-by-4.0")
```

## Data fields

- `audio`: embedded lossless FLAC audio decoded by Hugging Face
- `text`: ASR reference transcript
- `id`: deterministic public identifier without a local path
- `language`: `ar-en`
- `split`: `train` or `test`
- `data_source` and `source_dataset`: source identity
- `license` and `license_note`: row-level license status
- `domain`, `text_source`, `source_type`, `prompt_file`, `text_generation_model`, `tts_model`, and `tts_language`: generation provenance

## Data creation

All 9,617 files contain mono 16-bit PCM audio synthesized at 24 kHz with `gemini-2.5-pro-preview-tts`, using the `ar-EG` language setting and prebuilt voices. They are distributed in lossless FLAC format to reduce transfer and storage size without changing the samples. The text has two origins:

- 5,673 cleaned Egyptian Arabic-English lines adapted from ArE-CSTD.
- 3,944 domain-oriented lines generated from structured prompts for engineering, finance, justice, and medicine; the recorded text-generation model is `gemini-3.1-pro-preview`.

The corpus contains approximately 14.69 hours. Filter `license == "cc-by-4.0"` to obtain the 3,944-row author-created subset without duplicating its audio in a second configuration.

## Component licenses

This repository uses `license: other` because its components have different licenses. The `license` and `license_note` fields identify the terms for each row.

- **Prompt-generated rows:** Abdelrahman R. Hashem releases the author-created prompts, generated text, and resulting synthetic audio under **CC BY 4.0**. [Google's Gemini API terms](https://ai.google.dev/gemini-api/terms) state that Google does not claim ownership of generated content.
- **ArE-CSTD-derived rows:** the [ArE-CSTD card](https://huggingface.co/datasets/SDAIANCAI/Ar-En-Code-Switching-Textual-Dataset) body states **CC BY-NC-SA 4.0**, although its Hugging Face metadata currently says CC BY-SA 4.0. This release follows the stricter card-body statement. Attribution and share-alike obligations apply, and commercial use is not permitted.

This section records provenance and risks; it is not legal advice.

## Intended uses and limitations

- Training or evaluating Arabic-English code-switched ASR.
- Studying synthetic speech augmentation.
- The synthetic test partition is not evidence of real-world human-speech robustness.
- TTS voices do not represent real demographic, acoustic, or prosodic diversity.
- TTS artifacts and prompt repetition can make results easier or more model-specific than human speech.

## Companion human review annotations

Review decisions, correction deltas, and canonical split membership for the upstream human dataset are prepared as a separate metadata-only repository:

[`abdo1819/arabic-english-code-switching-review-annotations`](https://huggingface.co/datasets/abdo1819/arabic-english-code-switching-review-annotations)

That package contains no human audio and joins to the pinned upstream dataset by row index.

## Citation

```bibtex
@dataset{hashem2026syntheticcodeswitchasr,
  author    = {Hashem, Abdelrahman R.},
  title     = {Synthetic Arabic-English Code-Switched Speech for ASR},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/abdo1819/arabic-english-code-switching-synthetic-asr}
}
```

Please also cite ArE-CSTD when using the `synthetic` configuration.

## Contact

Abdelrahman R. Hashem — `arh13@fayoum.edu.eg`