Voice Similarity Tool

A simple tool to determine which TTS voice sounds most similar to a reference audio sample. It normalizes audio files, computes voice embeddings using Resemblyzer, and ranks similarity using cosine distance. Everything runs inside Docker with a minimal CLI wrapper script.

Features

• Neural voice embeddings (Resemblyzer)
• Automatic audio preprocessing (24 kHz, mono, loudness-normalized)
• Fully containerized environment
• Simple usage: bash run.sh old.wav new*.wav
• Works on Windows using Git Bash

Requirements

• Docker Desktop
• Git Bash (Windows) or any Bash shell
• ffmpeg installed on the host
• WAV files (one reference + multiple candidates)

Setup

docker build -t voice-similarity .

Usage

bash run.sh old.wav new*.wav

The script will:

1. Convert and normalize all audio files into _normalized_audio/
2. Run the Docker container
3. Print a similarity ranking (highest = most similar)

Example output:

Most similar voices:
voice17.wav: similarity = 0.8123
voice05.wav: similarity = 0.7991
...

File Overview

Dockerfile                    # Python + ffmpeg environment
find_most_similar_voice.py    # Embedding + similarity computation
run.sh                        # Audio preprocessing + Docker execution

Windows Notes

Git Bash may rewrite POSIX-style paths. To avoid this, run.sh can use:

WINDOWS_WORKDIR="/$(pwd)"
LINUX_WORKDIR="//data"

These ensure volume mounting works correctly on Windows.

License

MIT