Fish Audio Voice Design: Designing a Character or Brand Voice From a Text Prompt

Every video needs a voice, and the default voices are the ones audiences recognize as stock. Fish Audio voice design solves that differently than cloning does. Instead of uploading a recording of someone, you write a description of the voice you want ("a calm woman in her thirties, warm and slightly dry, unhurried pace") and the model generates candidate voices that match. You preview, keep what works, and lock the winner as a reusable voice ID for narration, dialogue, and video workflows.

That distinction matters when you are building a cast rather than duplicating a person. A brand mascot, a recurring explainer host, a game NPC roster, or a narrator for a faceless channel: none of these have real source recordings to clone from. Voice design creates the persona from scratch. If you have a real voice to preserve, use cloning instead, and our Fish Audio voice clone page covers that path. This guide is about the invented voice: how to brief it, audition it, lock it, and feed it into a production pipeline.

Why a designed voice is worth the effort

Consistency is the product. A brand voice that changes between videos reads as amateur even when each individual voice is good. Once you design a voice and save it, every future script in TTS, every talking-head scene that accepts reference audio, and every ad variant uses the same timbre, pacing, and character. The voice becomes an asset you own, like a logo, except it took ten minutes to design.

There is a second reason to design rather than pick from a catalog: catalog voices are shared. The popular narrator voices are sitting in thousands of other people's videos. A designed voice is unlikely to be reused by anyone else because it came out of your specific description plus your specific audition choices.

Step 1: Write the voice description like a casting brief

The Fish Audio voice design tool on Moosky takes a natural language prompt. Treat it like the casting note you would hand a voice director. The parts that actually steer the result:

• Age range and gender presentation. "Man in his 50s," "young woman, twenties." Age changes texture more than almost any other single word.
• Character and tone. Warm, cold, dry, eager, weary, playful, authoritative. Name the attitude, not just the sound. A "friendly" voice is vague; "a neighbor who actually wants to help you" is a brief.
• Pace and rhythm. Unhurried, clipped, conversational, deliberate. Pacing is what separates a voice that reads as trustworthy from one that reads as salesy.
• Accent and diction if they matter. Regional accents, transatlantic clarity, casual mumble. Skip them if neutral is fine; specifying an accent you don't need constrains the model.
• Use context. Mentioning "narrator for documentary" or "upbeat host for product ads" nudges the register because those phrases carry training associations.

One caution: write the voice you want, not a famous person. Describing a celebrity is both a rights problem and an unreliable instruction. Describe attributes instead, they generalize better and they are legal.

Step 2: Preview with real lines, not filler

The tool generates a short sample you can listen to immediately. The single biggest mistake at this stage is auditioning with generic demo text. Preview with a line your actual content would speak: your hook, your tagline, the first sentence of a typical script. A voice that sounds intriguing reading "this is a sample" can be unlistenable on your specific sentence, because pacing and emphasis habits surface against real copy.

Run three previews minimum, each with a different candidate voice, each reading the same real line. Same material, side by side, is the only audition method that gives you a trustworthy comparison. Expect the first description to miss. Narrow it like you would narrow a prompt for an image: too fast, add "measured pace"; too bright, add "lower register, subdued"; too flat, name an emotion the persona would feel about the subject.

Step 3: Lock the voice you chose

When a preview matches, save it. What you get is a preview audio file plus a persistent Fish Audio voice model ID. That ID is the important half. It is reusable, it goes into any Fish Audio TTS call on Moosky, and it does not expire just because a new model ships. The preview audio itself is also useful, more on that below.

Record the ID and the description that produced it in your project notes. Six months from now, when you are casting the next character or replacing a voice after a rebrand, the brief plus the ID is the entire documentation of the decision. This is the step most people skip, and it is the one that makes the voice a real brand asset.

Step 4: Use it in narration and video

The locked ID plugs into two pipelines.

Text to speech. Feed the ID into Fish Audio TTS and synthesize narration scripts directly: voiceover, explainer tracks, audiobook-style readings for faceless YouTube videos. Billing is per use on your Moosky credit balance, with the cost shown before you generate, so long scripts are predictable.

Video. Modern lip-sync video workflows are audio driven: you bring the speech and the model animates the face to it. That means the designed voice carries into talking-head and UGC scenes. Generate the line in TTS with your voice ID, pass the audio into the video generation as speech input, and the on-screen character speaks in your brand's voice. If you are building a recurring creator persona, this is how the face from your character-consistency workflow gets a matching sound. Our talking-head UGC workflow guide covers the video side of that handoff.

Where designs go wrong

Three failure modes come up repeatedly. First, over-briefing: a description with eight adjectives produces mush because nothing dominates. Two or three strong attributes beat a paragraph. Second, auditioning on wrong material, covered above, but worth repeating because it is the most common reason a voice sounds great in preview and wrong in production. Third, not re-listening after lock: render one full script through the actual TTS pipeline before committing the voice to a campaign. The designed sample is short-form; some voices drift on longer paragraphs, and you want to catch that at asset time, not at launch time.

And remember the boundary with cloning: if the brief changes from "I want a character who sounds like X" to "I want Sarah to sound like herself in every video," stop designing and clone Sarah's actual samples from the voice clone page instead. Design invents; cloning preserves. If you want the fastest sample-to-speech path, CosyVoice voice cloning on Moosky enrolls a voice from a 5–20 second clip for free and turns any script into speech in that voice.

The workflow in one paragraph

Write a casting brief with age, character, pace, and context. Generate candidate voices, preview each against a real line from your content, and narrow by editing the description. Save the winner and record its voice model ID alongside the brief. Use the ID in Fish Audio TTS for narration and pass that audio into lip-synced video scenes for on-screen work. Re-test one full script before launch. That is the entire loop: description, audition, lock, reuse. The voice design tool is on Moosky AI with pay-per-use credits, no subscription.