Clone a voice with Enhanced Text-to-Speech

Enhanced Text-to-Speech can build a custom voice from a short reference recording — a single clip of someone speaking — and then narrate your manuscript in that voice. The clone is created on your computer, with no long training wait. This is a zero-shot clone: you don't supply hours of audio or wait for a model to retrain. You give Arqenne one short sample, and it does the rest.

Everything happens locally. The reference audio you provide and the cloned voice it produces stay on your computer and are never uploaded.

Voice cloning is part of Enhanced Text-to-Speech, the studio-quality engine, on the Creator (one clone) and Studio (unlimited clones) plans. It runs on macOS and Linux: on a Mac it uses the Apple Silicon GPU; on Linux it runs on the processor, or on an NVIDIA graphics card with the NVIDIA edition. For an overview of the built-in voices and the two engines, see Narrate your manuscript with Text-to-Speech.

Before you start

You need:

1. A Creator or Studio plan. Enhanced Text-to-Speech and voice cloning aren't available on Free or Starter — see pricing.
2. At least 12 GB of memory. On a computer with less, Arqenne explains why instead of offering the model. 16 GB is recommended.
3. The Enhanced Text-to-Speech model downloaded. Download it once from Settings › Models. It's about 3.5 GB on macOS and 2.1 GB on Linux. After that, cloning works offline. See Downloading and managing local models for where models live.
4. A reference recording. A WAV file of 5 to 15 seconds (up to 50 MB) of the voice you want to clone. Record it with any recorder app, save it as WAV, then import it into Arqenne.

Permission first. Only clone voices you have the right to use — your own voice, or someone who has clearly agreed to it. A cloned voice is a likeness; treat it like one.

How cloning works

The flow has three parts: you create a speaker and give it a reference recording, Arqenne builds the cloned voice from that recording, and then you assign the speaker to paragraphs and generate audio like any other voice.

1. Create a new speaker. In the manuscript editor, switch to Audio mode (using the Text / Audio toggle). Each paragraph becomes its own lane with a voice selector. Create a new speaker for your clone and give it a name you'll recognize.

2. Give it a reference recording. Click Choose WAV file and pick your clip, or choose a reference you've used before from the saved list. This recording is the entire basis for the clone, so its quality matters more than anything else (see Recording a good reference below).

3. Let Arqenne build the voice. From your reference, Arqenne builds the cloned voice on-device. There's no lengthy training step — the voice is ready to use right away.

4. Assign the speaker and generate. In Audio mode, set a paragraph's lane to your cloned speaker, then use Generate Audio to render that paragraph. Use Play All to play every rendered paragraph in order; the current paragraph is highlighted as it plays. As with any voice, edited paragraphs are flagged stale so you know to regenerate just those.

Recording a good reference

The clone can only be as good as the clip you give it. A few minutes spent on a clean reference recording pays off in every paragraph you generate afterward.

A clean 10-second clip almost always beats a noisy longer one. When a clone sounds off, the reference is the first thing to improve.

Tuning the result

Once a clone is producing audio, you can adjust how it sounds without re-recording. Each speaker carries its own settings, so you can tune the clone independently of the built-in voices.

• Speed — how quickly the voice reads, from 0.5× to 2×. Nudge it slower for dramatic narration or faster for brisk dialogue.
• Audio effects — reverb (from a small room to a cathedral) and presence, to place the voice in a space.

Make a small change, regenerate a paragraph, and listen. Tuning is quick because generation is local, so you can iterate freely.

Privacy and responsible use

• Nothing is uploaded. Your reference recording and the cloned voice it produces stay on your computer. Cloning runs entirely on-device, and your text and audio never leave the machine.
• Generated audio is yours. Clones produce the same 24 kHz WAV files as any other voice, saved alongside your project.
• Only clone voices you're allowed to. Clone your own voice, or a voice whose owner has clearly agreed. Don't impersonate real people without permission.

Troubleshooting

The clone sounds off, robotic, or unlike the reference

Improve the reference recording first — it's the single biggest factor. Use a cleaner clip: less background noise, one speaker only, even tone and pace. If your reference is very short, try one closer to 10–15 seconds. Then point the speaker at the better recording and regenerate.

Voice cloning isn't available

Check three things: your plan is Creator or Studio, your computer has at least 12 GB of memory, and the Enhanced Text-to-Speech model is downloaded in Settings › Models (its card says which of these is missing). The Standard Text-to-Speech voices work in the meantime.

The first clone is taking a long time

Arqenne loads the Enhanced Text-to-Speech model into memory the first time you use it, so the first paragraph takes longer than the rest. On Linux, the NVIDIA edition is many times faster than the Standard edition for cloned voices — see Arqenne on Linux.

Where the model lives

To confirm the Enhanced Text-to-Speech model is present, look in:

Manage or delete it from Settings › Models rather than by hand.

See also

• Narrate your manuscript with Text-to-Speech — the Audio editor, lanes, and the built-in voices
• Downloading and managing local models — what downloads, how big, and where it's stored
• AI writing suggestions — on-device help while you draft