---
pretty_name: Vedavani
tags:
  - automatic-speech-recognition
  - speech
  - sanskrit
  - vedic
  - low-resource
  - whisper
  - audio
language:
  - sa
license: apache-2.0
task_categories:
  - automatic-speech-recognition
multilinguality: monolingual
annotations_creators:
  - expert-generated
source_datasets: []
paper: https://arxiv.org/pdf/2506.00145v1
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: transcription
      dtype: string
  splits:
    - name: train
      num_examples: 24623
    - name: validation
      num_examples: 3078
    - name: test
      num_examples: 3078
  download_size: ~5.4GB
  dataset_size: ~
---

# Vedavani: A Benchmark Corpus for ASR on Vedic Sanskrit Poetry

**Vedavani** is the first benchmark dataset for **automatic speech recognition (ASR)** on **Vedic Sanskrit poetry**, consisting of richly annotated verses from the *Rig Veda* and *Atharva Veda*. This corpus captures the unique **prosodic structure**, **phonetic complexity**, and **chanting style** found in traditional Vedic recitation.

🔗 **Paper**: [Vedavani: A Benchmark Corpus for ASR on Vedic Sanskrit Poetry (ACL 2025)](https://arxiv.org/pdf/2506.00145v1)  
📁 **GitHub Repository**: [https://github.com/SujeetNlp/Vedavani](https://github.com/SujeetNlp/Vedavani)  
📜 **License**: Apache License 2.0


---

## 📦 Dataset Contents

This repository contains:

- `train.csv` — Metadata for training set  
- `validation.csv` — Metadata for validation set  
- `test.csv` — Metadata for test set  
- `Audio_files` — Audio files in WAV format (segmented and aligned) [Due to hugging face restrictions files are organized in folder containing maximum 9000 files each. While using them in training/testing kindly move all the files in one single directory.]
- `README` — Textual documentation  

Each CSV includes:
- `path`: Relative path to audio file
- `transcription`: Ground-truth text in **Devanagari script**, including **prosodic markers**

---

## 📊 Dataset Statistics

| Property             | Value                    |
|----------------------|--------------------------|
| Total Duration       | ~54 hours                |
| Total Samples        | 30,779                   |
| Verses from Rig Veda | 20,782                   |
| Verses from Atharva Veda | 9,997               |
| Avg. Audio Length    | 6.36 seconds             |
| Vocabulary Size      | 64,082 unique words      |

### Data Splits

| Split      | # Samples |
|------------|-----------|
| Train      | 24,623    |
| Validation | 3,078     |
| Test       | 3,078     |

---

## Use Cases

Vedavani is particularly useful for:

- Fine-tuning and benchmarking ASR models (e.g., Whisper, IndicWhisper, Wav2Vec2)
- Studying phonetic alignment in Sanskrit poetry
- Low-resource speech processing
- Prosody-aware speech models

---


# Citation
```bibtex
@article{
  title={Vedavani: A Benchmark Corpus for ASR on Vedic Sanskrit Poetry},
  author={Sujeet Kumar, Pretam Ray, Abhinay Beerukuri, Shrey Kamoji, Manoj Balaji Jagadeeshan, and Pawan Goyal},
  journal={https://arxiv.org/pdf/2506.00145v1},
  year={2025}
}
```