---
pretty_name: SaUrSiSynth
language:
  - skr
  - ur
  - sd
license: cc-by-4.0
task_categories:
  - translation
  - automatic-speech-recognition
  - text-to-speech
tags:
  - speech-translation
  - synthetic-speech
  - indo-aryan
  - low-resource
  - saraiki
  - urdu
  - sindhi
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    default: true
    drop_labels: true
    data_files:
      - split: train
        path: Train_data/*
      - split: validation
        path: Dev_data/*
      - split: test
        path: Test_data/*
---

# SaUrSiSynth

Tri-parallel synthetic spoken language translation corpus for **Saraiki**, **Urdu**, and **Sindhi**.

## Summary

| Item | Value |
|------|------:|
| Aligned triplets | 2,687 |
| Audio files | 8,061 |
| Total duration | 6.91 h |
| Saraiki / Urdu / Sindhi | 2.31 / 2.30 / 2.31 h |
| Domains | Gen | Cul | HLT |
| Splits | Train 6,447 | Dev 807 | Test 807 |
| Sample rate | 16 kHz mono WAV |

**Pipeline:** Saraiki-Urdu bitext -> NLLB-200-distilled-600M completes Sindhi (**Urdu->Sindhi** pivot only) -> cleaning -> OmniVoice TTS -> quality filter -> domain packaging (Gen / Cul / HLT).

## Layout

```
huggingface/
  Train_data/   # train WAVs
  Dev_data/     # validation WAVs
  Test_data/    # test WAVs
  parallel_text/
    triplets.tsv
    sample_manifest.jsonl
    SAMPLE.csv
  manifest.jsonl
  metadata.csv
  dataset_meta.json
  load_saursisynth.py
  README.md
```

AudioFolder uses root `metadata.csv` (`file_name` = relative WAV path) plus the split folders above.

## Schema

`manifest.jsonl` / `metadata.csv` fields:

| Field | Description |
|-------|-------------|
| `utt_id` / `id` | Utterance id |
| `triplet_id` | Shared id across Sa/Ur/Si |
| `split` | Train / Dev / Test |
| `lang` / `language` | Saraiki / Urdu / Sindhi |
| `lang_code` | Sa / Ur / Si |
| `domain` | Gen / Cul / HLT |
| `gender` | Speaker prompt gender |
| `text` | Transcript |
| `wav_path` / `file_name` | Relative path to WAV |
| `duration_sec` | Clip duration |

## How to load

```python
from datasets import load_dataset
ds = load_dataset("hjav/saursisynth")
print(ds)
# DatasetDict with train / validation / test; audio decoded via AudioFolder
```

Local helper (same folder layout):

```python
from load_saursisynth import load_with_datasets, load_manifest
ds = load_with_datasets(".")
rows = load_manifest(".")
```

## License

**CC-BY-4.0** for the dataset release (confirm before public upload if institutional review is required).

Code for reproducing the pipeline is Apache-2.0 on GitHub.

## Citation

```bibtex
@inproceedings{saursisynth2027,
  title={{SaUrSiSynth}: A Multilingual Synthetic Corpus for Low-Resource Spoken Language Translation in Indo-Aryan Languages},
  author={Javed, },
  booktitle={Proc.\ },
  year={2027}
}
```

## Links

- GitHub (code): https://github.com/hjy895/SaUrSiSynth
- Dataset: https://huggingface.co/datasets/hjav/saursisynth