How to clone your own voice locally with VoiceStudio

Start with a ten-second project

A useful first voice clone is a short recording you can judge. Record your own voice, ask it to say one new sentence, and compare the result with the reference. Once that works, save the voice for a longer project.

This guide uses the local desktop app. The website's prerecorded examples are listening samples; you make the new audio in VoiceStudio on your computer. You do not need a VoiceStudio Cloud account.

These fully local steps assume a platform with a supported local runtime. The v0.5.6 Intel Mac desktop build requires a remote backend and does not provide local synthesis on that Mac. Check the platform notes before installing.

• The app: install the build for your platform from Download and complete its local-runtime setup
• A reference: 5–15 seconds of your own clean speech, or an adult speaker's recording with clear permission for the intended use
• A cloning engine: an installed engine that accepts reference audio, with its required model files downloaded
• A small test: one or two sentences and headphones for checking the result

1. Choose a cloning engine that fits your computer

In the app, open Settings → Models and check the TTS engine. An engine being listed does not mean its runtime and weights are ready. Use the installation controls offered for your machine, then wait for the model to become available.

OmniVoice is the default engine in this release. It supports reference-based cloning on CPU and compatible accelerators, but a CPU run can take substantially longer. The engine guide recommends at least 6 GB of dedicated GPU memory for CUDA or ROCm. That figure is a recommendation for that engine, not a universal VoiceStudio requirement or a guarantee that a project will fit. Apple Silicon uses unified memory and is a separate case.

MOSS-TTS-Nano is another option for supported languages when you want a lighter, reference-only cloning engine. Its guide lists CPU and CUDA operation. The managed installer is not offered on Intel Macs in v0.5.6. It does not support voice-design instructions or speed control, so those settings are not a way to improve its clone. Check the MOSS engine requirements before selecting it.

KittenTTS is useful for preset-voice speech, but it does not clone a reference voice. Choosing a preset-only engine will not complete this exercise. Use the model library to compare capabilities, then trust the app's availability state for your actual installation.

Initial installation and model downloads need a connection. With local engines and their files installed, synthesis can run locally. Configuring a remote engine or other external service changes that data path; “local” is a choice about the engines you use, not a promise about every optional integration.

2. Record a reference worth copying

Open Voice cloning. Use Choose a voice for a first reference, or Change voice if a voice is already selected. Choose Upload audio or Record.

For a first try, record the following in a quiet room at a comfortable speaking volume. This is an example reference, not a required consent statement:

Every small project starts with a clear idea. Today I'm recording a short sample in my normal voice, with enough space to finish each sentence.

Keep the microphone position steady. Aim for one speaker, no background music, little room echo, and no clipped peaks. Leave a short natural pause between sentences, but remove a long silent lead-in. A short clear sample is easier to diagnose than a noisy interview or an overprocessed recording.

Listen to the uploaded or recorded sample before continuing. If it sounds distant, distorted, or noisy, replace it now. Those problems can carry into the generated voice.

Under Optional details → Transcript, enter the words actually spoken in the reference when your engine needs them. This field describes the sample; the new words belong in Script. If you use OmniVoice without a reference transcript, an installed speech recognizer may be needed. Follow any ASR readiness message rather than assuming another model will download automatically.

Keep this exercise within 5–15 seconds. Voice Clone accepts longer clips up to 75 seconds, but engines handle them differently. In v0.5.6, OmniVoice can select a speech-rich passage from references longer than 20 seconds. A manually supplied whole-clip transcript cannot describe that selected passage reliably. To remove that ambiguity, trim the audio and transcript to the same short passage. The release's reference-length table explains the other engine limits.

3. Generate one new sentence

If the app asks you to save the new reference, give it a clear name such as “My voice · quiet room”, or select Use without saving for this test. Saving a profile makes the reference reusable; it is not a model-training step.

1. Enter this new text in Script: “The next train leaves at nine thirty. Please check the platform before you go.”
2. Choose English as the output language for this example
3. Leave Production overrides at their defaults for the first take
4. Select Synthesize audio and watch the loading or generation status

The language setting guides synthesis. It does not translate the script. If you later want another language, write or obtain the words in that language and choose an engine that supports it.

The first generation may include model loading or downloads, so do not treat its total wait as the engine's steady-state speed. Avoid launching several renders while a model is still being prepared.

4. Compare the take with the reference

Play Latest take, then replay the reference. Similarity alone is not enough. Check four things before you turn the test into a longer script:

• Identity: does it keep the speaker's broad tone, rather than switching to a different-sounding voice?
• Words: are “nine thirty” and every other word spoken once, in the right order?
• Pacing: is the pause between sentences useful, and is the final word complete?
• Audio: is there clipping, hiss, echo, an abrupt cut, or a distracting breath?

If a result misses, change one thing at a time. First try a cleaner reference; next check the transcript; then simplify the script. Keep the sentence unchanged while comparing references so you know what caused the difference. Spell out a number or abbreviation when its intended pronunciation is ambiguous.

Use Save as… on a take you want to keep. Open the saved audio outside VoiceStudio and play it from beginning to end. A successful generation is not yet an editorial approval.

5. Save the useful reference and protect it

Select Save as voice profile if you have not saved the reference yet. Use a name that distinguishes the speaker and recording session. Keep your original reference, permission record, and approved output together in your own project files.

For your own saved profile, Edit voice → Voice ownership offers Record consent statement. This records an ownership declaration. Treat it as a record of what you declared, not independent proof of identity or a substitute for someone else's permission. Local synthesis does not require this verification in v0.5.6.

The release also lets you replace a saved reference through Edit voice → Replace reference. Replacing it clears the profile's locked take and own-voice verification because the new audio might be a different speaker. Recheck the result after a replacement instead of assuming the saved voice is unchanged.

A profile and its reference can enable future synthesis. Protect them like other project source material, and do not share the raw recording just because the generated clip is approved for publication.

When the first take does not work

Synthesize audio is unavailable

Check that there is a reference, a non-empty Script, and a ready cloning-capable TTS engine. Read the inline blocker. A model name in the list is not proof that its dependencies and weights are installed.

The voice sounds wrong

Check that you selected the intended saved voice. Listen for another speaker, music, or strong echo in the reference. Try a new short sample before changing advanced controls. Some combinations of language, accent, and model will work better than others.

The reference transcript is rejected

For OmniVoice references longer than 20 seconds, trim the sample and its transcript to one matching passage, or follow the engine's automatic passage-selection path. Do not paste the whole transcript beside a shorter reference.

Generation is slow or runs out of memory

Distinguish a progressing model download from synthesis. Try a shorter Script, close other heavy applications, and use an engine suitable for your hardware. A higher time budget does not add RAM or GPU memory.

Sources and version notes

Reviewed on 4 October 2026 against the published Electron v0.5.6 source and engine guides. The example sentences are original exercises. This article does not claim a speed benchmark, a measured voice-similarity score, or native testing on every operating system.

• v0.5.6 release and shipped reference-replacement changes
• Released reference and profile controls, voice ownership controls
• OmniVoice requirements and reference behavior, application and model license boundary

For the shorter reference page, see Clone a voice. For an overview of where saved voices can be used, see the voice-cloning workflow.