Voice Cloning Training Data Length Evaluation

A test collection for subjectively evaluating how different lengths of training data affect voice cloning quality in one-shot workflows.

Background

One-shot voice cloning refers to voice synthesis systems that clone a voice from a single audio sample (as opposed to multi-speaker training datasets). This repository explores how the duration of that single sample affects output quality.

Test Methodology

Source Samples

The voice-samples/ directory contains reference audio at three durations:

File	Duration	Purpose
general-1-min.wav	~1 minute	Minimal training data baseline
general-3-mins.wav	~3 minutes	Moderate training data
general-5-mins.wav	~5 minutes	Extended training data

Evaluation Process

1. Clone voice using each of the three source samples
2. Generate TTS output using identical text prompts across all three clones
3. Subjectively compare output quality across dimensions:
   • Voice similarity / timbre accuracy
   • Naturalness and prosody
   • Handling of varied content types (conversational, technical, emotional, etc.)

Text Samples

The text-samples/ directory contains varied prompts designed to test different aspects of TTS performance.

Output Structure

Generated voice clones will be stored in output/ organized by source sample duration.