Clone Your First Voice

Before you begin, make sure you have the following:

• GPU with 8+ GB VRAM — NVIDIA (CUDA 11.8+), AMD (ROCm 5.7+), or Apple Silicon
• Microphone — or a pre-recorded WAV/MP3 file of the target voice
• Thumper-Run installed — download here if you haven’t already
• ~4 GB free disk space — for the OpenVoice model weights and dependencies

OpenVoice is available in the Thumper catalog. One-click install handles all dependencies.

1. Open Thumper-Run and navigate to the Catalog
2. Search for OpenVoice or browse the Audio category
3. Click Install — the runtime downloads the model weights (~2.5 GB) and sets up the Python environment
4. Wait for the status to change to "Ready" (typically 3–5 minutes)

You should see output similar to this in the console:

# Console output during installation
[1/4] Cloning OpenVoice repository...
[2/4] Creating virtual environment...
[3/4] Installing dependencies (torch, torchaudio, openvoice)...
[4/4] Downloading model weights...
  ✓ OpenVoice v2 base model (1.2 GB)
  ✓ OpenVoice v2 converter (1.3 GB)
✓ Installation complete. Ready to launch.

To verify the installation, click Launch and wait for the health check to pass. The status indicator turns green when the service is running.

Alternative: Qwen3-TTS

If you prefer to design a voice from a text description rather than providing an audio sample, consider Qwen3-TTS instead. Qwen3-TTS lets you describe a voice style in natural language (e.g., "a warm, deep male voice with a slight British accent") and generates speech matching that description. It’s available in the same Audio category in the catalog.

Comparing OpenVoice and Qwen3-TTS

Disk Space Requirements

The quality of your voice clone depends heavily on the quality of the reference recording. Here’s how to get the best results.

Recording Requirements

• Duration — 10–30 seconds of continuous speech
• Format — WAV or MP3 (WAV preferred for lossless quality)
• Sample rate — 16 kHz or higher (44.1 kHz recommended)
• Channels — mono preferred, stereo is accepted
• Environment — quiet room, no background noise or echo

Good vs Bad Recording Conditions

If you don’t have a microphone, you can use any existing clean audio recording. Podcast clips, audiobook excerpts, or voice memos work well as long as they meet the quality criteria above.

Preparing Your Audio File

If your recording needs cleanup before use, follow these steps:

1. Trim silence from the beginning and end of the file
2. Remove any sections with coughs, pauses longer than 2 seconds, or background interruptions
3. Normalize the audio level to -3 dB peak to avoid clipping
4. Export as 16-bit WAV at 44.1 kHz sample rate

Free tools like Audacity can handle all of these steps. If you’re using a phone recording, transfer the file to your computer and convert from M4A/OGG to WAV before uploading.

# Convert audio formats with ffmpeg (optional)
ffmpeg -i recording.m4a -ar 44100 -ac 1 -acodec pcm_s16le reference.wav
 
# Trim to first 30 seconds
ffmpeg -i reference.wav -t 30 -c copy reference_trimmed.wav

With OpenVoice running and your reference audio ready, you can generate cloned speech in a few clicks.

Step-by-Step

1. In the OpenVoice interface, click Upload Reference and select your WAV or MP3 file
2. The model analyzes the voice characteristics (tone, pitch, cadence) — this takes 2–5 seconds
3. Enter the text you want spoken in the Text Input field
4. Select the output language (defaults to English)
5. Click Generate
6. The generated audio plays automatically and saves to your output directory

Understanding the Output

After generation completes, you’ll see a playback widget with the generated audio. The output file is saved as a WAV in your Thumper output directory (typically ~/.local/share/tr-desktop/outputs/openvoice/). Key things to listen for:

• Tone match — the generated voice should have the same pitch range and timbre as the reference
• Cadence — speech rhythm and pacing should feel natural, not robotic
• Pronunciation — words should be clearly articulated without slurring or skipping
• Background — output should be clean with no artifacts, hiss, or echo

If the first result isn’t perfect, try regenerating — slight variations occur between runs. You can also adjust speed and style settings (covered in the next section) to fine-tune the output.

Cross-Lingual Cloning

OpenVoice supports cross-lingual voice cloning: you can provide an English voice sample and generate speech in Chinese, French, or any other supported language. The cloned voice characteristics transfer across languages.

Cross-lingual cloning preserves the speaker’s unique voice timbre (the quality that makes a voice recognizable) while adapting pronunciation to the target language. The accent will be a blend — not a perfect native accent in the target language, but clearly the same speaker.

Example Text Inputs

Use these sample texts to test different speech patterns and verify clone quality:

# Declarative statement
The quick brown fox jumps over the lazy dog near the riverbank.
 
# Question with rising intonation
Have you ever wondered what it would be like to travel to Mars?
 
# Exclamation with emphasis
What an incredible performance that was! I have never seen anything like it.
 
# Technical content with numbers
The server processes approximately 2.5 million requests per second at peak load.
 
# Conversational with pauses
Well, I think... honestly, it depends on what you are looking for.