File size: 1,634 Bytes
c4e0faa
c97d780
 
 
 
 
 
 
 
 
c4e0faa
c97d780
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
---
license: cc-by-nc-sa-4.0
language:
  - ar
  - en
task_categories:
  - automatic-speech-recognition
pretty_name: Clean One-Speaker STT (AR dialectal + MSA + EN)
size_categories:
  - 10K<n<100K
---

# Clean_One_Speaker_STT_Split_EN_AR

Mixed STT dataset — Saudi dialectal Arabic + Modern Standard Arabic + English —
with train/validation/test splits, built for fine-tuning multilingual ASR (e.g. Whisper)
without catastrophic forgetting of English.

## Composition

- **Dialectal Arabic (~72%)** — [Sebssihakim/Clean_One_Speaker_SADA_Split](https://huggingface.co/datasets/Sebssihakim/Clean_One_Speaker_SADA_Split)
  (cleaned single-speaker SADA; SDAIA / Saudi Broadcasting Authority, CC BY-NC-SA 4.0)
- **MSA (~8%)** — FLEURS `ar_eg`, full official splits (CC-BY; read MSA), plus the
  MSA rows already present in SADA
- **English (~17%)** — LibriSpeech clean (CC-BY), sampled from its official splits
  (train.100 / validation / test); transcripts lowercased

## Columns

`audio` (16 kHz), `text`, `language` (`ar`/`en`), `dialect`
(SADA dialect name, `ModernStandardArabic`, or `English`)

## Split methodology

SADA partitions reused as-is (stratified by dialect, seed 42, Maghrebi removed).
FLEURS and LibriSpeech samples come from their **own** official splits (seed 42),
so no cross-split leakage is possible. SADA splits are segment-level: the same
show/speaker may appear in more than one split.

## License & attribution

SADA is CC BY-NC-SA 4.0 (SDAIA & Saudi Broadcasting Authority); this derivative keeps
the same license. LibriSpeech (Panayotov et al., 2015) and FLEURS (Conneau et al., 2022)
are CC-BY 4.0.