Tamil Voice Dataset (tamilVoice2)

A curated dataset of Tamil speech audio recordings paired with text transcriptions, designed for training and evaluating Automatic Speech Recognition (ASR) systems for the Tamil language.

Dataset Description

This dataset contains Tamil voice recordings and their corresponding transcriptions, collected for speech-to-text model training. The audio data is stored in compressed archive format and includes both the raw audio chunks and the aligned transcription files.

Use Cases

• Automatic Speech Recognition (ASR): Train Tamil-language speech-to-text models
• Voice Interface Development: Build voice-enabled applications for Tamil speakers
• Language Model Fine-tuning: Improve acoustic models for Dravidian languages
• Linguistic Research: Study Tamil phonetics and spoken language patterns

File Inventory

File	Description	Size
audio_chunks.rar	Compressed archive of Tamil speech audio recordings	~1.0 GB
audio_trans.rar	Compressed archive of text transcriptions aligned to audio chunks	~1.0 GB
.gitattributes	Git LFS configuration for large file storage	—
README.md	Dataset documentation (this file)	—

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("edwixx/tamilVoice2", trust_remote_code=True)

Access Control

This dataset is gated and requires manual approval. To request access:

1. Go to https://huggingface.co/datasets/edwixx/tamilVoice2
2. Click "Request Access" and provide your use case
3. Access will be granted manually

Extracting Data

The audio and transcription data are stored as RAR archives. Use unrar or a Python RAR library to extract:

# Install unrar
sudo apt-get install unrar

# Extract audio chunks
unrar x audio_chunks.rar

# Extract transcriptions
unrar x audio_trans.rar

Metadata

• Author: Anurag Kanade
• Language: Tamil (ta)
• License: MIT
• Hosted by: Hugging Face Datasets
• Created: 2024-04-28
• Last Updated: 2026-06-25
• Storage Size: ~2.15 GB
• Downloads: 2

Citation

@misc{edwixx-tamilVoice2,
  author = {Anurag Kanade},
  title = {tamilVoice2},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/datasets/edwixx/tamilVoice2}}
}