🗣️ Synthetic Voice Description Dataset for Voice Conversion

Overview

This dataset contains 3.180 synthetically generated voice samples, each paired with a caption that roughly describes how the voice sounds. Captions typically include information about the age, gender, accent, and sometimes general vocal qualities (e.g., "old woman, Irish accent").

All voices are synthetically generated and do not represent any real person. This makes the dataset ideal for training and evaluating voice conversion models without infringing on personhood or identity rights.

📁 Dataset Contents

A CSV file containing pairs of:
- description: A short text caption describing the voice
- matched_filename: The name of the corresponding .wav file

Example rows from the CSV:

description,matched_filename
"old woman, Irish accent",00399-000.wav
"old man, valley girl accent",01074-001.wav
"young man, Scottish accent",18297-001.wav
"middle-aged man, Chinese accent",00952-002.wav

A TAR archive that contains:
- The voice-captions.csv file (as described above)
- All 3.180 corresponding .wav audio files

🎯 Purpose

The dataset is designed to serve as a target files for voice conversion tasks. It can be used to:

Convert existing voice data into synthetic voices with known, described properties

⚙️ Google Colab for Seed-VC Fine-Tuning

A community-developed fine-tuning notebook for Seed-VC (a voice conversion model) is available on Google Colab:

🔗 Seed-VC Fine-Tune Notebook

This notebook allows you to take real or synthetic voice data and convert it into one of the target voices.