README
Egyptian Arabic ASR Clean 72 h
Dataset Summary
This corpus contains ≈72 h of Egyptian‑Arabic speech aligned to text.
Audio has been resampled to 16 kHz mono WAV, transcripts are normalised Arabic (no diacritics, Tatweel, digits verbalised), and the data are split 80 / 10 / 10 into train / validation / test.
Supported Tasks and Leaderboards
| Task | Tags | Notes |
|---|---|---|
| Automatic Speech Recognition | asr, speech-recognition | Primary use‑case |
| Forced Alignment / VAD | alignment, vad | Clips ≤ 25 s |
Languages
The dataset is predominantly Egyptian Arabic (ar‑EG).
~85 % of recorded hours are male speakers; speaker IDs are unavailable.
Dataset Structure
Data Fields
| Field | Type | Description |
|---|---|---|
audio | Audio | Pointer to WAV @ 16 kHz |
text | string | Normalised Arabic transcript |
duration | float | Seconds (post‑resample) |
dataset_source | string | One‑letter code A–D |
Splits
| Split | Hours |
|---|---|
| train | ≈57.6 |
| validation | ≈7.2 |
| test | ≈7.2 |
Source Data
| Code | Raw Hours | Description |
|---|---|---|
| A | ~465 | Long clips, heavy overlap |
| B | ~65 | Similar to A, shorter |
| C | ~5 | Dual‑channel conversations |
| D | ~2.5 | YouTube excerpts |
| Other | <5 | Minor sources |
Only ≈12 % of the original 570 h survived the cleaning pipeline.
Data Collection and Processing
- Format Unification – convert all audio to 16 kHz WAV.
- Deduplication – drop exact audio/text duplicates; remove nulls.
- Metadata Pruning – retain only core fields.
- Text Normalisation – strip diacritics, Tatweel, punctuation, Latin letters; verbalise digits; fix common glyph errors; run CAMeL‑Tools morphology checks.
- Alignment Diagnostics – compute chars/s and words/s; flag extreme values.
- Duration Filtering – keep clips 0.5–25 s.
- Shuffle & Split – 80 / 10 / 10 random split, uploaded as
datasets.DatasetDict.
Usage Example
from datasets import load_dataset
ds = load_dataset("your-username/egyptian_arabic_asr_clean72h")
print(ds["train"][0]["audio"].sampling_rate) # 16000
print(ds["train"][0]["text"])
Similar repos
Linked in READMEs
README
Egyptian Arabic ASR Clean 72 h
Dataset Summary
This corpus contains ≈72 h of Egyptian‑Arabic speech aligned to text.
Audio has been resampled to 16 kHz mono WAV, transcripts are normalised Arabic (no diacritics, Tatweel, digits verbalised), and the data are split 80 / 10 / 10 into train / validation / test.
Supported Tasks and Leaderboards
| Task | Tags | Notes |
|---|---|---|
| Automatic Speech Recognition | asr, speech-recognition | Primary use‑case |
| Forced Alignment / VAD | alignment, vad | Clips ≤ 25 s |
Languages
The dataset is predominantly Egyptian Arabic (ar‑EG).
~85 % of recorded hours are male speakers; speaker IDs are unavailable.
Dataset Structure
Data Fields
| Field | Type | Description |
|---|---|---|
audio | Audio | Pointer to WAV @ 16 kHz |
text | string | Normalised Arabic transcript |
duration | float | Seconds (post‑resample) |
dataset_source | string | One‑letter code A–D |
Splits
| Split | Hours |
|---|---|
| train | ≈57.6 |
| validation | ≈7.2 |
| test | ≈7.2 |
Source Data
| Code | Raw Hours | Description |
|---|---|---|
| A | ~465 | Long clips, heavy overlap |
| B | ~65 | Similar to A, shorter |
| C | ~5 | Dual‑channel conversations |
| D | ~2.5 | YouTube excerpts |
| Other | <5 | Minor sources |
Only ≈12 % of the original 570 h survived the cleaning pipeline.
Data Collection and Processing
- Format Unification – convert all audio to 16 kHz WAV.
- Deduplication – drop exact audio/text duplicates; remove nulls.
- Metadata Pruning – retain only core fields.
- Text Normalisation – strip diacritics, Tatweel, punctuation, Latin letters; verbalise digits; fix common glyph errors; run CAMeL‑Tools morphology checks.
- Alignment Diagnostics – compute chars/s and words/s; flag extreme values.
- Duration Filtering – keep clips 0.5–25 s.
- Shuffle & Split – 80 / 10 / 10 random split, uploaded as
datasets.DatasetDict.
Usage Example
from datasets import load_dataset
ds = load_dataset("your-username/egyptian_arabic_asr_clean72h")
print(ds["train"][0]["audio"].sampling_rate) # 16000
print(ds["train"][0]["text"])