RonitMehta260704/kh-en-dataset

KYNMAW — Khasi Bhashini Multi-Task Dataset

The largest open multi-task speech + text dataset for Khasi, an Austroasiatic language spoken by roughly 1.5 million people in Meghalaya, Northeast India — built to help bring Khasi into modern speech recognition, text-to-speech, and machine translation systems as part of

If you're working on low-resource ASR, endangered-language NLP, Indic speech translation, or Northeast Indian languages, this dataset is for you.

---

---

Why This Dataset Matters

Khasi has no official status in India's 8th Schedule and is largely absent from existing Indic speech/NLP corpora — most Indic datasets (including Bhashini's own reference corpora) cover the 22 scheduled languages, which Khasi is not part of. That makes real, spoken-broadcast Khasi audio like this genuinely scarce as training data. This dataset exists to close that gap using authentic broadcast speech — not synthetic or read-aloud recordings — from All India Radio Shillong.

Dataset Summary

Config	Task	Input	Output	Status
asr_khasi (default)	Speech Recognition	Khasi audio (16kHz WAV)	Khasi transcript	🟡 audio ready, transcripts pending annotation
asr_english	Speech Recognition	English audio (16kHz WAV)	English transcript	🟡 audio ready, transcripts pending annotation
speech_translation	Speech Translation	English audio	Khasi text	🟡 audio ready, translations pending annotation
machine_translation	Machine Translation	English text	Khasi text	🟡 pending — verify upload path before use
movie_backtranslated	Legacy / pilot MT pairs	Movie audio clips	Khasi text (unofficial MT)	⚠️ pilot data, see caveat below

Note on movie_backtranslated: this config is left over from an earlier pilot pipeline that scraped subtitled Khasi-language films and back-translated English subtitles using an unofficial free MT endpoint. It is not human-verified and should be treated as noisy pilot data, not gold-standard translation — kept here for transparency, not held out as equivalent in quality to the AIR-Shillong-sourced configs above.

Quickstart

from datasets import load_dataset

# Default config — Khasi ASR
ds_asr_khasi = load_dataset("RonitMehta260704/kh-en-dataset")
print(ds_asr_khasi["train"][0])

# English ASR
ds_asr_en = load_dataset("RonitMehta260704/kh-en-dataset", "asr_english")

# Speech translation (English audio -> Khasi text)
ds_st = load_dataset("RonitMehta260704/kh-en-dataset", "speech_translation")

# Machine translation (English text -> Khasi text)
# NOTE: verify this config's data files exist at the path declared in this
# card's YAML before relying on it — it was missing from the repo metadata
# as of the last card revision and has been added here as a placeholder.
ds_mt = load_dataset("RonitMehta260704/kh-en-dataset", "machine_translation")

# Legacy pilot config (movie-derived, machine-translated — see caveat above)
ds_legacy = load_dataset("RonitMehta260704/kh-en-dataset", "movie_backtranslated")

Dataset Structure

Schemas below are approximate per config; the movie_backtranslated schema is confirmed from the repo's auto-generated dataset_info, the others reflect the intended pipeline output — regenerate dataset_info automatically (via push_to_hub) rather than hand-editing it, so byte counts and features always match what's actually uploaded.

asr_khasi / asr_english

{'audio': {'array': [...], 'sampling_rate': 16000}, 'duration_sec': 8.3, 'transcript': ''}

speech_translation

{'audio_english': {...}, 'audio_khasi': {...}, 'translation_en2kha': ''}

movie_backtranslated (confirmed schema, 399 rows)

{
  'audio': {'array': [...], 'sampling_rate': 16000},
  'transcript_khasi': str,
  'transcript_english': str,
  'duration': float,
  'source_video_id': str,
  'source_type': 'movie_backtranslated',
}

Data Collection Methodology (KYNMAW Pipeline)

AIR Shillong Archive (250 pages)
        │
        ▼
  [Scrape] English 0830 + Khasi 0745 bulletins per date
        │
        ▼
  [Process] Noise Removal → Silence Trimming → 16kHz Resampling
        │
        ▼
  [Segment] 2–20 second clips via silence detection (ffmpeg silencedetect)
        │
        ▼
  [Pair] English ↔ Khasi matched by broadcast date
        │
        ▼
  [Upload] 4 Bhashini-compatible HF dataset configs

🤝 How to Contribute

The audio backbone is here — the transcripts are the missing piece, and that's exactly where outside contributors can have outsized impact:

• Native Khasi speakers: transcribe or verify short clips (even 10 minutes of annotation helps — see the Contributing guide / open a PR).
• ASR practitioners: run Whisper or another multilingual model as a bootstrap pass on the English-audio configs, then have it human-corrected.
• Everyone else: star the repo, share it with anyone working on Northeast Indian languages, low-resource ASR, or Indic NLP — visibility drives contributors, contributors drive transcripts.

Open an issue or discussion on this repo's Community tab to get involved.

Limitations & Ethical Considerations

• Transcripts are placeholders. transcript_khasi / transcript_english fields in the ASR/ST configs are currently empty pending annotation (see above) — this dataset is audio-complete but text-incomplete today.
• Source rights: audio is scraped from AIR Shillong's public broadcast archive via Prasar Bharati. Confirm and document your redistribution permissions for this content — government-broadcaster material can carry its own reuse terms distinct from the CC BY 4.0 license applied to this derived dataset.
• Content neutrality: AIR news/current-affairs bulletins may incidentally include government messaging or political content. If you're building a general-purpose ASR corpus, consider flagging or filtering such segments so the dataset doesn't skew toward one speaker register.
• movie_backtranslated quality: see the caveat in the summary table above — this config's Khasi text is unverified machine translation, not gold data.

Roadmap

• Native-speaker annotation pass on asr_khasi (highest priority)
• English transcript pass on asr_english
• Verify and populate machine_translation config data files
• Confirm/redocument AIR Shillong content redistribution terms
• Publish held-out test splits for benchmarking

License

This dataset is released under CC BY 4.0 — commercial and non-commercial reuse is permitted with attribution.

An earlier version of this card additionally described the dataset as "non-commercial research only," which conflicts with the CC BY 4.0 license declared in the metadata. If you intend to restrict commercial use, change the license: field to cc-by-nc-4.0 instead of leaving prose and metadata disagreeing — automated tools and Hub search filters read the metadata field, not the prose.

Source: Prasar Bharati / All India Radio (public broadcaster).

Compatible Models (KYNMAW Stack)

Task	Model
ASR	wav2vec 2.0 / XLS-R
TTS	FastSpeech 2 + HiFi-GAN
MT	IndicTrans2 / mBART fine-tune
ST	Seq2Seq with cross-lingual encoder

Citation

@dataset{kynmaw_khasi_bhashini_2025,
  author    = {unknown},
  title     = {KYNMAW Khasi  Multi-Task Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/RonitMehta260704/kh-en-dataset}
}