Quick Start

This guide gets you from a fresh install to your first generated clip. It starts with Kokoro, a small model that's fast even on CPU, and then moves on to voice cloning.

1. Launch OpenVoiceLab

./scripts/run.sh     # scripts\run.bat on Windows

Open http://localhost:7860. You'll see five tabs: Generate, Voices, Models, Data and Training. The footer shows the devices OpenVoiceLab detected (CUDA, MPS or CPU).

2. Load a model

On the Generate tab:

1. Set Model family to Kokoro.
2. Leave Checkpoint and Device (auto) as they are.
3. Click Load model.

The first time, OpenVoiceLab installs Kokoro's environment and downloads its weights. This takes a minute or two. Later loads take a few seconds.

3. Generate

1. Type something in Text:

   Hello! This is my first test with OpenVoiceLab.

2. Under Voice, choose Preset voice and pick a voice from Preset, such as af_heart. Kokoro voice names start with a language and gender code: af_ is American female, bm_ is British male, and so on.
3. Click Generate.

The clip plays when it's ready, along with a line such as 3.6s of audio in 0.8s (4.5× realtime). Every clip is saved in outputs/, and the History panel lists the clips from this session.

4. Clone a voice

Now try a model that can clone a voice from a reference clip.

1. Set Model family to Qwen3-TTS and choose the checkpoint 0.6B Base (voice cloning).
   • For other options, see Supported Models. VoxCPM2, Chatterbox Turbo and Echo-TTS are all good cloning models.
2. Click Load model. Loading the model frees the memory used by the previous one.
3. Under Voice, choose Voice library and pick one of the bundled voices, such as Alice. Or choose Upload clip and use your own 5-20 second recording.
4. Click Generate.

Some models also need the transcript of the reference clip. OpenVoiceLab fills it in with Whisper when it's missing, so you usually don't have to type it.

5. Try a dialogue

Load VibeVoice (1.5B), set Speakers to 2, give each speaker a voice, and enter:

Speaker 1: Welcome back to the show! Today we're talking about open-source speech models.
Speaker 2: Thanks for having me. It's been a big year for TTS.
Speaker 1: It really has. Let's dive in.

Each Speaker N: line uses that speaker's voice.

Next steps

• Generating Speech: every option on the Generate tab
• Managing Voices: build your own voice library
• Supported Models: what each model is good at