Getting Started

OpenVoiceLab is a local web app for running, comparing and fine-tuning open text-to-speech (TTS) models. It runs on your own computer: your text and voices never leave it.

Choose your path

Try it out (10 minutes)

1. Install OpenVoiceLab
2. Generate your first clip with a small, fast model
3. Try a cloning model with one of the bundled voices, or your own

Clone a voice (no training)

1. Record or find 5-20 seconds of clean speech
2. Add it on the Voices tab
3. Use it with any cloning model on the Generate tab

Fine-tune a custom voice (hours)

1. Collect 30+ minutes of one speaker's audio
2. Build a dataset
3. Train a VibeVoice LoRA adapter (needs an NVIDIA GPU)
4. Load the adapter on the Generate tab

Ways to choose a voice

Different models support different ways to pick a voice. The Generate tab only shows the options that the loaded model supports.

Option	What it does	Models
Preset voice	Built-in voices that ship with the model	Qwen3-TTS CustomVoice, Kitten TTS, Pocket TTS, Kokoro, VibeVoice Realtime
Voice library / Upload clip	Zero-shot cloning from a short reference clip	VibeVoice, Qwen3-TTS Base, VoxCPM, Chatterbox, OmniVoice, Echo-TTS, MOSS-TTS, Pocket TTS
Describe a voice	Voice design from a text description, e.g. "a warm, older British man"	Qwen3-TTS VoiceDesign, VoxCPM2, OmniVoice, MOSS VoiceGenerator
Default voice	Let the model pick a voice	Models that can speak without a reference
Fine-tuning	Train the model on 30+ minutes of a voice for the best likeness	VibeVoice

Some models also take a style instruction (for example "Very happy" or "whispering"). Several models can do multi-speaker dialogue: write the script as Speaker 1: ... lines and choose a voice for each speaker.

Cloning vs fine-tuning

Zero-shot cloning needs no training. You give the model a few seconds of reference audio and it imitates the voice right away. Modern cloning models (VoxCPM2, MOSS-TTS, Echo-TTS, Qwen3-TTS Base) are good enough that most people never need to fine-tune.

Fine-tuning trains a LoRA adapter on 30+ minutes of audio. It takes hours and an NVIDIA GPU, but captures a voice's style and consistency better over long content. OpenVoiceLab supports fine-tuning for VibeVoice.

Requirements

• OS: Linux, macOS or Windows
• Python: 3.10+, or uv, which installs Python for you
• Hardware:
  • Kitten TTS, Pocket TTS and Kokoro run fast on any CPU.
  • Most other models want a GPU: an NVIDIA GPU or an Apple Silicon Mac. Memory needed ranges from about 2 GB (Chatterbox Nano, MOSS-TTS Nano) to 24 GB+ (MOSS-TTS 8B, VibeVoice 7B).
  • Fine-tuning needs an NVIDIA GPU with 16 GB+ VRAM.
• Disk: a few GB for the app and model environments, plus the model weights (25 MB to 24 GB per checkpoint)

Next steps

1. Installation: set up OpenVoiceLab
2. Quick Start: generate your first clip
3. Supported Models: pick the right model for the job

Questions? See the FAQ, Troubleshooting, or ask on Discord.