toiar/Khasi_ASR_Dataset

Khasi ASR Dataset

The Khasi ASR Dataset is a large-scale speech recognition dataset for the Khasi language, an Indigenous language spoken primarily in Meghalaya, India. The dataset contains paired audio recordings and transcriptions designed for training and evaluating Automatic Speech Recognition (ASR) systems.

This dataset consists of 73,900 audio-transcription pairs with a total duration of approximately 101 hours, 19 minutes, and 54.36 seconds of speech data.

The dataset contains a mixture of real human speech recordings and synthetic speech generated using Khasi-OmniVoice-TTS-2. Additionally, a portion of the dataset contains Khasi-English code-switched speech, reflecting natural language usage commonly found in everyday conversations, education, media, and digital communication.

The dataset aims to support research and development of speech technologies for Khasi, including:

• Automatic Speech Recognition (ASR)
• Speech-to-Text Systems
• Language Modeling
• Speech Understanding
• Low-Resource Language Research
• Indigenous Language Preservation

Dataset Details

Metric	Value
Total Samples	73,900
Total Audio Duration	101h 19m 54.36s
Language	Khasi (kha)
Task	Automatic Speech Recognition
Format	Audio + Text
Audio Types	Real + Synthetic
Synthetic Speech Model	Khasi-OmniVoice-TTS-2
Code-Switching	Khasi-English

Data Structure

Each record contains:

{
    "audio": <audio_file>,
    "text": "Khasi transcription"
}

Features

Feature	Type	Description
audio	Audio	Speech recording (real or synthetic)
text	String	Corresponding Khasi transcription

Language

The dataset contains speech and text in Khasi, a member of the Austroasiatic language family spoken in Meghalaya, India. Some samples include Khasi-English code-switching, reflecting contemporary language usage among Khasi speakers.

Intended Uses

This dataset is suitable for:

• Training ASR models for Khasi
• Fine-tuning multilingual speech models
• Benchmarking low-resource ASR systems
• Speech technology research
• Academic research on Indigenous languages
• Language preservation initiatives

Supported model families include:

• Whisper
• NVIDIA NeMo ASR
• Wav2Vec 2.0
• MMS (Massively Multilingual Speech)
• SeamlessM4T
• Other speech foundation models

Data Collection

The dataset was curated from a combination of:

1. Real Khasi speech recordings paired with transcriptions.
2. Synthetic speech generated using Khasi-OmniVoice-TTS-2 from Khasi text corpora.

The inclusion of synthetic speech was intended to increase vocabulary coverage, improve data diversity, and expand training resources for low-resource ASR development.

Citation

If you use this dataset in your research, please cite:

@dataset{toiar_khasi_asr_dataset,
  author = {Toiarbor Mawlieg},
  title = {Khasi ASR Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/toiar/Khasi_ASR_Dataset}
}