VoxCPM2 Voice Design Guide: Write a Voice Like Casting Notes

Voice cloning copies a voice that already exists. Voice design creates one that does not. You describe the voice you want and the model builds it from scratch, with no recording, no microphone, and no reference clip.

VoxCPM2 designs voices from a plain-language description, and what it rewards is not length but coverage. Where other models give you a set of dropdowns or want a compact sketch, VoxCPM2 reads one continuous description that covers who the voice is, what it sounds like, and how it performs. The most useful way to think about it: you are not writing a prompt, you are writing casting notes.

Hear What You Can Design

Four voices, four descriptions. Every one below was generated from the exact description shown next to it, with no reference audio and no recording. Press play to hear each, then use the copy button to drop the description straight into Voice Design.

Notice the pattern in every description above: who the voice belongs to (a middle-aged man, a young boy), what it is made of (booming and resonant, thin and strained), and how and where it performs (slowly with long pauses, perfect for an epic movie trailer). Those three layers are the whole format, and the rest of this guide teaches you to write them.

The Three Layers

Every strong VoxCPM2 description stacks three layers. Each answers a different question, and skipping one is the most common reason a designed voice comes out generic.

Assembled, those three layers read as one continuous description:

A quiet raspy, elderly woman of a low-pitched voice with a distinct, grainy
texture and subtle breathy tremors. Delivers a slow tone at a very low volume,
perfect for historical narration.

Notice that the layers are not labelled or separated in the final text. VoxCPM2 takes ordinary prose, so the layers are a checklist for you, not a syntax for the model. Write naturally, then check that all three are present before you generate.

Layer 1: Identity

Identity is the shortest layer and the one people most often over-think. Gender, rough age, and a role are usually enough: "a middle-aged male broadcaster," "an elderly woman," "a young woman," "a man in his thirties."

A role does a lot of work here, because it carries assumptions the model can act on. "A broadcaster" implies clarity, projection, and pacing. "A drill sergeant" implies volume and clipped delivery. If a single word gets you most of the way to the character, use it and spend your detail budget on the other two layers.

Layer 2: Texture

Texture is what makes a designed voice sound like a specific person instead of a category. This is where the description earns its length. Useful vocabulary falls into a few groups:

• Pitch: low-pitched, deep, baritone, high-pitched, bright.
• Grain and quality: raspy, gravelly, grainy, smooth, silky, hoarse, rough, thin, rich.
• Air and resonance: breathy, airy, resonant, warm, nasal, hollow, booming.
• Character: magnetic, weathered, youthful, brittle, velvety, commanding.

Stack two or three of these rather than one. "A raspy voice" is a category, while "quiet and raspy, low-pitched, with a distinct grainy texture and subtle breathy tremors" is a person. The qualifiers matter as much as the adjectives: "subtle," "distinct," "slight," and "heavy" tell VoxCPM2 how far to push each trait.

Layer 3: Delivery

Delivery covers performance: pace, volume, emotion, and the setting the voice belongs to. "Delivers a slow tone at a very low volume." "Speaks quickly with rising urgency." "Measured and deliberate, with long pauses."

Emotion words carry the most weight here. Of everything in a description, the emotional vocabulary shifts the result the hardest. "Warm, slightly melancholic, slow" and "warm, enthusiastic, upbeat" produce noticeably different voices even when the rest of the description is identical. If a voice is close but the feel is wrong, change the emotion words before you touch anything else.

The setting is the other powerful part of this layer, and it deserves its own section.

Scene-Defined Voices

Naming where the voice belongs shapes the entire performance. "Perfect for historical narration," "perfect for an epic movie trailer," "suitable for late-night radio," "for a bedtime story."

This works because a setting bundles dozens of delivery decisions into a phrase everyone already understands. A movie trailer voice is not just deep, it is deliberate, heavily paused, and building. You could try to specify each of those, or you could name the trailer and let the model work backwards from it.

That makes the scene a genuine shortcut when you know the destination but cannot find the adjectives. If you can hear the voice in your head but cannot describe its texture, describe the room it is speaking in instead:

You can use both, of course. The strongest descriptions carry explicit texture and a named setting, so the model has the raw material and the context to apply it.

Match Your Text to the Emotion

VoxCPM2 does not only read your description, it reads the words you ask it to speak, and it infers prosody from them. This catches people out constantly.

A furious voice description paired with "The quarterly report is now available for download" will underdeliver, because the sentence gives the voice nothing to be furious about. The description and the script are both inputs, and they need to agree.

So when you are auditioning a designed voice, give it something worth performing:

• For a panicked character, use fragments, questions, and interruptions.
• For a trailer voice, use short declarative lines with hard stops.
• For a calm narrator, use flowing descriptive prose.
• For fury, give it something to be angry about.

The samples at the top of this guide were built that way. The trailer voice reads short lines with hard stops so the model turns each full stop into a dramatic pause, while the panicked boy gets fragments and a repeated question that escalate on their own. Same principle in both: the script does half the acting.

If a designed voice sounds flat, check the script before you rewrite the description. The problem is often the text.

Expect Variation, Then Save the Winner

Designed voices vary between generations, even with an identical description. This is by design, and it means the workflow is a loop rather than a single shot.

1. Write the description with all three layers present.
2. Generate two or three times on the same description. Do not change the wording yet, because you are sampling the range it produces, not testing a hypothesis.
3. Listen for direction, not perfection. Is the age right but the pace wrong? Is the texture right but too bright?
4. Edit one layer and repeat. Changing texture and delivery at once makes it impossible to hear which edit did what.
5. Save the winner. When a generation nails it, save that voice to your library so you never have to chase it again.

That last step is what turns a lucky generation into a reusable asset. Once a designed voice is saved, you can use it across any text with the timbre locked in, the same way you would use a cloned voice. Design once, then reuse it for a whole audiobook or series.

Start from a preset when you can. Voice Creator Pro ships starter presets for VoxCPM2, from broadcaster and documentary narrator to extreme emotions like fury, terror, and cold menace. Editing a description that already works is faster than writing one from scratch, and the presets double as worked examples of the layered format.

Design, Cloning, and Prompting Are Different Things

These three get conflated constantly, so to be clear:

• Voice design (this guide): build a new voice from a description. No recording needed.
• Voice cloning: copy a specific real voice from a short reference clip (3 to 10 seconds). VoxCPM2 does this too, with zero-shot cloning that scores among the top open models for speaker similarity on the standard benchmarks. Longer reference audio does not produce a better clone.

VoxCPM2 blurs the first two productively. A common workflow is to design a voice, save it to your library, and then use that saved voice as you would a cloned one.

One thing to know: a voice description is a design tool, not a cloning tool. When you clone a voice from a reference clip, VoxCPM2 takes its identity and texture from the recording, so a description written in the design box does not steer that clone. Design a voice or clone one, and treat the description as belonging to the first.

The other models in Voice Creator Pro design voices too, each in its own way. For how they compare, and which to reach for when, see the AI voice design guide, the OmniVoice voice design guide, the Qwen3 TTS prompting guide, and the DramaBox prompting guide.

Design a Voice in Voice Creator Pro

Voice Creator Pro runs VoxCPM2 with layered voice design, cloning, and 48kHz output, plus starter presets so you can begin from a working description and edit rather than write from scratch.

Designed voices are not limited to one-off clips. They work inside projects alongside cloned voices, so you can build a designed cast, give each character its own description, and run a full chapter or dialogue in one place.

The desktop app runs VoxCPM2 on your own machine, offline and with unlimited generations, as a one-time purchase on Windows and Mac. It needs around 6 GB of VRAM, which most recent graphics cards have. Or try it free in your browser, with no GPU and nothing to install.