RonitMehta260704/kh-en-dataset

KYNMAW — Khasi Bhashini Multi-Task Dataset

Team: ITerative Bytes | Hackathon: Bhashini Khasi Language Model Training
Roles: Ronit Satish Mehta (MLOps), Vipul Mhatre (ASR/TTS), Varun Vyas (NLP/MT),
Ankush Pandey (Annotation), Saniyaa Shetty (AI Architect)

---

Dataset Description

Multi-task audio + text dataset for Khasi (kha), a low-resource Austroasiatic
language of Meghalaya, Northeast India. Sourced from AIR Shillong (All India Radio)
broadcasts via Prasar Bharati's public archive.

Built to support all 4 Bhashini model training tasks:

Config	Task	Input	Output
asr_khasi	Speech Recognition	Khasi audio (16kHz WAV)	Khasi transcript
asr_english	Speech Recognition	English audio (16kHz WAV)	English transcript
speech_translation	Speech Translation	English audio	Khasi text
machine_translation	Machine Translation	English text	Khasi text

---

Usage

from datasets import load_dataset

# Task 1 & 2 — ASR / TTS (Khasi)
ds_asr = load_dataset("RonitMehta260704/kh-en-dataset", "asr_khasi")
print(ds_asr["train"][0])
# {'audio': {'array': [...], 'sampling_rate': 16000},
#   'duration_sec': 8.3, 'transcript_khasi': '', ...}

# Task 3 — Machine Translation
ds_mt = load_dataset("RonitMehta260704/kh-en-dataset", "machine_translation")
print(ds_mt["train"][0])
# {'transcript_english': '...', 'transcript_khasi': '...', 'date_key': '...'}

# Task 4 — Speech Translation
ds_st = load_dataset("RonitMehta260704/kh-en-dataset", "speech_translation")
print(ds_st["train"][0])
# {'audio_english': {...}, 'audio_khasi': {...}, 'translation_en2kha': ''}

---

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

---

Annotation Status

Transcripts are empty by default. Fill via:

• Native Khasi speaker annotation (recommended)
• Whisper pseudo-labelling (English only, as bootstrap)
• Bhashini ULCA crowdsourcing portal

---

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    = {ITerative Bytes},
  title     = {KYNMAW Khasi Bhashini Multi-Task Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/RonitMehta260704/kh-en-dataset}
}

License

CC BY 4.0 — non-commercial linguistic research only.
Source: Prasar Bharati / All India Radio (public broadcaster).