Voice Cloning
Clone any voice from a recording, audio file, or YouTube video and use it for text-to-speech in the Lab's TTS tab.
Voice Cloning
Voice Cloning lets you replicate any voice using a short audio sample. Provide a reference voice, type your text, and generate speech that sounds like the original speaker. No training or fine-tuning required.
Cloning happens in the TTS tab of the Lab. Your saved voices and the built-in voices live in one picker, so you can switch between a cloned voice and a ready-made one without changing tabs.
The TTS tab has three panels:
- Voice (left) - Pick a voice, or record, upload, or import a new one
- Text to speak (center) - Enter text, pick a model, and set the language
- Advanced settings (right) - Fine-tune how the model generates speech
Below the panels is a History section that stores all your past generations. A player bar along the bottom edge plays the latest result, with playback speed, volume, and a download button.
The first time you open the TTS tab, a short walkthrough points out each step. Click Show walkthrough (the button next to Generate) to see it again.
Add a Reference Voice
The first step in voice cloning is providing a reference audio sample. Use the buttons in the Voice panel:
Record
Click Record to record your voice directly. Click Stop when you are done. This is useful for quick tests or when you want to clone your own voice.
Tips for best results:
- Use a quiet environment with minimal background noise
- Speak naturally for 5 to 10 seconds
- Avoid long recordings. Shorter, clean samples produce better clones, faster
For a deep dive on what makes a good reference clip, read How to Pick the Right Reference Audio for Voice Cloning. For more on why short clips work better than long ones, see How Many Minutes of Audio Do You Need for Voice Cloning?.
Upload
Click Upload to select an audio file from your computer. WAV, MP3, OGG, FLAC, AAC, M4A, OPUS, AIFF, and WMA files all work.
Once the transcript is ready, a Save Voice window opens with the file name filled in. Change the name if you like and click Save to keep the clip in your library.
Import from YouTube
On the desktop app, click YouTube to open the import dialog.
- Paste a YouTube URL into the field
- Click Fetch to download the audio
- Pick a start time, preview the clip, then click Use this clip to load it as your reference voice
Clip Length
Aim for 5 to 10 seconds of clean speech. A badge next to the panel title shows the clip length and warns you when it is too short or too long for the current model.
If a clip is longer than 15 seconds, a Trim voice clip window opens. Drag the selection to the part you want and click Use trimmed clip. End the selection on a natural pause or at the end of a sentence. Cutting off mid-word hurts clone quality.
Transcript
Below the waveform is the Transcript. When you load a reference voice, the built-in speech-to-text model fills it in for you.
To fix a word the model got wrong, click Edit. A reminder appears that this box is only for fixing transcription mistakes. The words you want the voice to say go in the Text to speak box. Click Edit anyway, make your fix, then click Done.
Important: The transcript must match exactly what is being said in the audio. Even small mismatches reduce clone quality. If the transcript does not match the clip when you generate, you'll see This transcript doesn't match your audio. Click Fix to replace it with the words spoken in your audio.
The Voices Panel
Click the voice row at the top of the Voice panel to open Voices. Every voice you can use is in here.
- Your voices - Voices you recorded, uploaded, imported, or designed. Designed voices carry a Design badge.
- Top voices - Hand-picked premium voices
- Built-in voices - Ready-made speakers, shown when the selected model has them, such as Kokoro
Click the play button on any row to hear a sample, and click the name to use that voice. Use the pencil and trash icons on your own voices to rename or delete them.
Searching
Type in Search all voices... to search your voices, the built-in voices, and the community library all at once. Use the language menu next to the search box and the tag chips below it to narrow the results. Click More filters for the full tag list.
Community results show below your own matches. Click a community voice's name to add it to your library and start using it right away, or click the add button to save it to your library for later.
Voices imported from Voice Search also appear under Your voices.
Saving a Voice
After loading a reference voice you are happy with, click Save Voice below the transcript to add it to your library. Give it a descriptive name so you can easily find it later.
Saved voices show up in the Voices panel, in Audiobooks, and in the Voice Changer.
Generating Speech
Text Input
In the Text to speak field, enter the text you want the cloned voice to say. The Lab is designed for short-form testing. Try different sentences and styles to dial in the perfect voice for your use case. A paragraph or two works well here.
Two additional tools are available in the text input area:
- Dictation (microphone icon) - Speak your text instead of typing it
- Expression tags (smiley icon) - Insert vocal expressions into your text
Numbers, dates, phone numbers, units, codes, and hashtags are read out the way you would say them.
Click Generate in the top bar when you are ready.
Long Text
Lab generations are capped at 200 words per request (200 characters on Chatterbox). If your text is longer, a This text is long window offers to Create a project. Projects break the text into smaller chunks for cleaner output. If your clip isn't saved yet, you'll be asked to name the voice first so the project can use it.
For full scripts or articles, use Audiobooks.
Pasting a Script With Speakers
When you paste a script that has two or more speaker names, an Assign voices to speakers window opens. Pick a voice for each speaker and click Create project to turn the script into a project with a voice for every speaker. Click Cancel to keep the pasted text in the Lab instead.
Expression Tags
Expression tags let you add natural vocal expressions like laughter, sighs, or surprise sounds at any point in your text. Click the smiley icon to see available tags.
Note: Expression tags are only supported by OmniVoice and by Chatterbox with Quality set to Lower. When you switch to a model that doesn't support them, the tags are removed from your text so they aren't read out loud.
Example usage:
So I was walking down the street [laughter] and you won't believe what happened next [surprise-oh]Emotion
When you clone with Qwen3, two extra menus appear under the text box. Pick an Emotion such as Happy, Sad, or Whisper, then set its Intensity from Very mild to Extreme. Leave Emotion on Neutral for a plain read.
Warnings Before Generating
Two checks run when you click Generate:
- Your text doesn't match the voice sample - Your text is written in a different alphabet than the reference transcript. Click Edit my text to retype it, or Generate anyway.
- Check the language setting - Your text looks like it is in a different language than the one selected. Click Switch to the suggested language or Keep current setting. If the model does not support that language at all, you'll see Language not supported with an option to switch to a model that does.
Model Selection
Click the model name next to Model, above the text box, to open Choose a model. Each model shows ratings for quality, speed, and price, plus the languages it supports.
| Model | Languages | Description |
|---|---|---|
| OmniVoice | 600+ | Fast, high-quality cloning with expression tag support |
| Qwen3 | 10 | Top-quality cloning with emotion control. Slower |
| VoxCPM 2 | 30 | Studio-quality 48kHz cloning |
| Chatterbox | 23 | Lighter cloning model with expression tags |
| Gepard 1.0 | 4 | Lightweight cloning that doesn't need a reference transcript |
Some models come in two sizes. When they do, a Quality switch appears next to the language menu. Lower is faster and Higher sounds better. For Chatterbox, Lower is the English-only Turbo model and Higher is the 23-language Multilingual model. On VCP Cloud, the switch is hidden for Qwen3.
Each model handles languages, accents, and speaking styles differently. Experiment with multiple models to find the best match for your use case.
Advanced Settings
The Advanced settings panel sits on the right of the TTS tab. On a narrower window, click the sliders icon at the end of the row under the text box to open it instead. A dot on the icon means a setting has been changed. Each model has its own settings. Click Reset to restore the defaults at any time.
OmniVoice Settings
| Setting | Default | Description |
|---|---|---|
| Speed | 0.8 | Speech speed multiplier. 1.0 is normal speed. |
| Length | 0 | Target length in seconds. Set to 0 to fit the text automatically. Desktop app only. |
| Quality | 32 | Higher improves quality but takes longer to generate. |
| Clarity | 2.0 | Higher makes speech clearer and more on-target. Lower is more relaxed but can drift or mumble. Too high can add artifacts. |
| Noise reduction | On | Cleans up background noise and artifacts for a clearer result. |
| Clean reference | On | Trims silence and evens out the reference audio before cloning. Turn off if your clip is already clean. |
| Polish output | On | Applies light cleanup and volume levelling to the result. Turn off for the raw model output. |
Qwen3 Settings
| Setting | Default | Description |
|---|---|---|
| Variation | 0.9 | Higher adds expressiveness and variety between runs. Lower stays consistent and predictable. |
| Boldness | 0.95 | Lower sticks to the safest sounds for a stable read. Higher lets bolder delivery through but can cause odd emphasis. |
| Candidates | 50 | How many candidate sounds are weighed at each step. Fewer is steadier, more allows extra variety. |
| Repetition control | 1.05 | Higher reduces stutters and loops. Too high can sound unnatural. |
VoxCPM 2 Settings
| Setting | Default | Description |
|---|---|---|
| Voice match | 2.0 | How closely the output matches your reference voice and text. Higher stays closer, lower is more varied. |
| Quality | 10 | Higher improves quality but takes longer to generate. |
| Seed | 0 | Use the same number to get the same result again. 0 picks a new random seed each run. |
Chatterbox Settings
| Setting | Default | Description |
|---|---|---|
| Expressiveness | 0.5 | How dramatic and emotional the delivery is. Higher is more expressive but can speed up speech. |
| Pacing | 0.5 | Higher speeds up the pacing for a more natural read. Lower slows down and stays closer to the reference voice. |
Gepard 1.0 Settings
| Setting | Default | Description |
|---|---|---|
| Variation | 0.4 | Higher adds variety between runs. Lower stays consistent and predictable. |
| Faithfulness | 1.0 | How closely the result follows your input. Higher is more faithful, lower allows more variation. |
| Candidates | 0 | How many candidate sounds are weighed at each step. Fewer is steadier, more allows extra variety. |
Use Cases
Content Creation
Clone a consistent narrator voice for YouTube videos, podcasts, or audiobooks. Record a short sample of the desired voice, save it to your library, then generate all your scripts using that voice.
Multilingual Content
Take a single voice sample and generate speech in 600+ languages with OmniVoice. Ideal for creating localized versions of video content, training materials, or product demos without needing native speakers.
Prototyping and Previews
Quickly prototype how dialogue will sound before hiring voice actors. Test different delivery styles, pacing, and tones by adjusting the advanced settings.
Accessibility
Generate natural-sounding audio versions of written content for visually impaired users. The expression tags add personality that makes listening more engaging than robotic TTS.
Social Media and Marketing
Clone a brand voice for consistent audio across social media posts, ads, and promotional videos. Save the voice to your library for reuse across campaigns.