README.md

---
license: cc-by-4.0
language:
- skj
task_categories:
- automatic-speech-recognition
- audio-classification
task_ids:
- keyword-spotting
- speaker-identification
- audio-language-identification
pipeline_tag: automatic-speech-recognition
pretty_name: Seke (SKJ) Language Corpus — Endangered Language Alliance
size_categories:
- n<1K
tags:
- endangered-language
- seke
- nepal
- sino-tibetan
- low-resource
- linguistics
- audio
- field-recording
- whisper
- wav2vec2
---

# 🏔️ Seke (SKJ) Language Dataset
## Endangered Language Alliance × Internet Archive

> **Seke** (`skj`) is a critically endangered Sino-Tibetan language spoken by approximately 700 people
> in the five villages of Upper Mustang district, Nepal, and in diaspora communities in New York City.
> This dataset represents one of the most complete public audio corpora of Seke ever assembled.
> Dataset created by Anil Tamang (himalaya-ai).
---

## 📊 Dataset Statistics

| Metric | Value |
|--------|-------|
| Total audio clips | 529 |
| Total duration | 17.8 minutes |
| Number of speakers | 2 |
| Sample rate | 16000 Hz |
| Format | 16-bit PCM WAV (mono) |
| Language ISO | `skj` |
| Date range | 2018–2021 |
| Assembled | 2026-04-25 |

### Quality Tier Distribution

| Tier | Count | Description |
|------|-------|-------------|
| A | 20 | Gold: elicited clips, controlled studio recording, short utterances (<10s), SNR>20dB |
| B | 205 | Silver: field recordings, good quality, longer utterances |
| C | 304 | Bronze: field recordings with background noise, or long sessions |

### Linguistic Type Distribution

| Language Description | 515 |
| Primary Text | 14 |


---

## 🔤 Language Information

- **ISO 639-3**: `skj`
- **Family**: Sino-Tibetan → Tibeto-Burman → Bodish
- **Dialect**: Primarily Chhusang (Tshugsang) dialect
- **Typology**: Verb-final word order, ergative alignment, evidential marking
- **Script**: No standardized orthography (romanized transcriptions in this dataset)
- **Endangerment**: Critically endangered (~700 speakers)
- **Region**: Upper Mustang, Gandaki Province, Nepal; NYC diaspora

---

## 🗂️ Dataset Fields

| Field | Type | Description |
|-------|------|-------------|
| `audio` | Audio | 16kHz mono WAV, trimmed and normalized |
| `duration_s` | float | Duration in seconds |
| `gloss_en` | string | English translation/gloss (from ELA elicitation) |
| `transcription_seke` | string | Whisper ASR output (romanized approximation) |
| `phonemes_ipa` | string | IPA phonemes from Allosaurus (where available) |
| `grammatical_categories` | string | Comma-separated grammatical category tags |
| `linguistic_type` | string | OLAC type: Lexicon / Language Description / Primary Text |
| `quality_tier` | string | A/B/C quality tier |
| `speaker_name` | string | Speaker name |
| `recording_date` | string | Session date |
| `source_identifier` | string | Internet Archive item identifier |
| `source_url` | string | Direct link to IA item |
| `is_swadesh` | bool | Part of Swadesh list elicitation |
| `swadesh_number` | int | Swadesh list number (-1 if not applicable) |
| `is_alternative_form` | bool | Alternative pronunciation of same sentence |
| `snr_db` | float | Signal-to-noise ratio in dB |
| `loudness_lufs` | float | Original LUFS loudness (pre-normalization) |

---

## 🚀 Usage

```python
from datasets import load_dataset

# Load full dataset
ds = load_dataset("Titung/seke-nepali-dataset")

# Load only high-quality elicited clips (Tier A)
ds_gold = load_dataset("Titung/seke-nepali-dataset", split="train")
ds_gold = ds_gold.filter(lambda x: x["quality_tier"] == "A")

# Fine-tune Whisper on Seke
from transformers import WhisperForConditionalGeneration, WhisperProcessor
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v3")
processor = WhisperProcessor.from_pretrained("openai/whisper-large-v3")

# Prepare features
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor(
        audio["array"], sampling_rate=audio["sampling_rate"],
        return_tensors="pt"
    ).input_features[0]
    batch["labels"] = processor.tokenizer(
        batch["gloss_en"]
    ).input_ids
    return batch

ds_prepared = ds["train"].map(prepare_dataset)
```

---

## 📖 Source & Attribution

All audio data sourced from the **Endangered Language Alliance (ELA)** archive on Internet Archive:
- Collection: https://archive.org/details/elalliance
- ELA website: https://www.elalliance.org
- Primary researcher: **Ross Perlin**, Co-Director, ELA
- Primary speakers: **Rasmina Gurung**, **Nora Gurung**, and others
- Recorded 2018–2021 at Columbia University and in Upper Mustang, Nepal
- Supported in part by the National Endowment for the Humanities

Original data archived under **Creative Commons Attribution 4.0** (CC-BY-4.0).
This dataset derivative is also released under **CC-BY-4.0**.
Please cite ELA and the original depositors in any academic use.

### Suggested Citation
```
Endangered Language Alliance. (2021). Seke (SKJ) Language Recordings.
Internet Archive. https://archive.org/details/elalliance

Dataset assembled by [Anil Tamang/himalaya-ai] (2026).
Available at: https://huggingface.co/datasets/Titung/seke-nepali-dataset
```

---

## ⚠️ Important Notes

1. **Transcriptions are approximate**: The `transcription_seke` field contains Whisper ASR output
   on an unseen language — treat as a phonetic approximation, not verified orthography.
   Manual validation by a Seke speaker is strongly recommended.
2. **No standardized orthography**: Seke has no official writing system. Romanizations may vary.
3. **Ethical use**: This data documents a critically endangered language. Please use it to
   support language preservation, not for commercial exploitation.
4. **Contact ELA**: For research partnerships or access to additional materials:
   https://www.elalliance.org/contact