Seif-Eldeen-Sameh/asr_codeswitched_dataset

Arabic/English Code-Switched ASR Dataset

Audio + transcripts of code-switched Egyptian Arabic and English speech, assembled to fine-tune ASR for Arab-world lecture content where dialectal Arabic and English technical vocabulary alternate within sentences.

Composition

Source	Description
EJUST custom recordings	Locally recorded/segmented code-switched clips (segments_codeswitched.csv + WAVs)
MohamedRashad/arabic-english-code-switching	~12,480 public code-switched clips (≤25 s each)

Total: 34.1 hours · 45,189 clips · avg 2.7 s/clip (utterance-level, ideal for Whisper-class ASR).

Schema

Column	Type	Description
audio	Audio(16kHz)	Mono waveform
transcript	string	Code-switched transcript (dialect + English terms preserved)

Load

from datasets import load_dataset, Audio
ds = load_dataset("Seif-Eldeen-Sameh/asr_codeswitched_dataset", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

Intended use

Fine-tuning ASR (e.g. Whisper) for code-switched Arabic/English. See the model trained on it: whisper-medium-arabic-codeswitched.

Limitations & ethics

• Egyptian dialect dominant; not representative of all Arabic dialects.
• Lecture/talk domain; technical vocabulary skew.
• Ensure you have rights to any audio you add; respect speaker consent.

License

MIT (custom portion). The public source dataset retains its own license — review before redistribution.