A strong Voice Design prompt is a casting brief, not a pile of flattering adjectives. It should tell a reviewer what job the voice must do, who will hear it, how the delivery should feel, and which traits would make the result wrong.
The fastest way to improve consistency is to separate three decisions: the voice brief, the audition copy, and the approval record. The brief defines the target. The audition reveals how each candidate behaves on real material. The approval record gives the team a reference when the project resumes, expands, or moves into another language.
Start with the job, not the sound
Before choosing attributes, write one sentence that describes the voice's responsibility:
A calm product guide who helps first-time users complete setup without sounding slow, theatrical, or patronizing.
That sentence is useful because it includes a role, a listener, an outcome, and boundaries. Compare it with "warm, professional, engaging voice." The second version sounds positive but gives two reviewers no reason to make the same choice.
Ask five questions before designing:
- Who is listening?
- What should they understand, feel, or do next?
- Where will they hear the voice?
- How long will they listen at one time?
- What would make the delivery feel wrong for this context?
A game character, product tutorial, audiobook narrator, and public announcement may all need clarity. They do not need the same rhythm, emotional range, or relationship with the listener.
Build a structured voice brief
Use attributes that can be heard and reviewed. A practical brief covers seven fields:
| Field | Question | Example |
|---|---|---|
| Role | What function does the voice serve? | Museum audio guide |
| Listener | Who is the voice addressing? | Curious adults walking through an exhibition |
| Delivery | How should each line arrive? | Conversational, assured, never performative |
| Pace | What rhythm supports the task? | Measured, with space around dates and names |
| Texture | What vocal quality fits the role? | Clear mid-range voice with a soft, dry texture |
| Emotional range | How much variation is appropriate? | Gentle curiosity with restrained emphasis |
| Avoid | Which traits are unacceptable? | No trailer voice, breathiness, sarcasm, or sing-song cadence |
The "avoid" field is especially valuable. Positive attributes often overlap. Exclusions draw a clearer boundary around the result.
Keep traits compatible. "Fast and meditative," "intimate stadium announcer," or "youthful with the authority of a retired judge" may force the generator and reviewer to guess which direction matters most. If the creative tension is intentional, explain the priority: "Fast enough for a short product tour, but with calm sentence endings."
Translate adjectives into audible behavior
Broad adjectives become useful when you define what they mean in the recording.
| Weak direction | Testable direction |
|---|---|
| Friendly | Speaks to one person, uses light sentence endings, avoids sales energy |
| Authoritative | Clear onset, steady pace, firm emphasis on actions, no booming delivery |
| Energetic | Quick transitions and bright emphasis, with clean consonants and no shouting |
| Natural | Small changes in rhythm, restrained emphasis, no repeated pitch pattern |
| Premium | Controlled pace, low vocal effort, precise articulation, minimal theatricality |
Do not describe clothing, lighting, camera angles, or a fictional studio. Those details can help an image prompt, but they do not give a voice reviewer an audible acceptance criterion.
Write audition copy that represents the real job
A voice can sound excellent on one polished sentence and fail on the material that matters. Test it with a short audition script that includes the difficult parts of the project.
For a product guide, include:
- a welcoming sentence
- a numbered action or sequence
- a product name or acronym
- a warning or limitation
- a longer explanatory sentence
- a calm confirmation at the end
For a character voice, include neutral dialogue, emotional pressure, a quiet response, a proper noun, and a line that must remain intelligible during action. For long-form narration, include dialogue, exposition, punctuation changes, and a paragraph long enough to reveal repetitive cadence.
Keep the audition script constant across candidates. If every voice reads different copy, reviewers cannot tell whether they prefer the voice or the sentence.
Use a scorecard instead of memory
Run a blind or lightly labeled review when practical. Score each candidate from one to five on the same criteria:
| Criterion | Review question |
|---|---|
| Role fit | Does this sound like the person the brief describes? |
| Clarity | Can a first-time listener understand every important word? |
| Pace | Does the rhythm support the task and listening duration? |
| Emotional control | Is the expression useful without becoming distracting? |
| Pronunciation | Are names, numbers, and specialist terms credible? |
| Listening fatigue | Would this remain comfortable across the intended duration? |
| Boundary check | Are any excluded traits present? |
Add one required note: "Why did this candidate win?" A score without a reason is hard to use later. A short decision such as "Candidate B stayed clear on instructions and never became promotional" becomes part of the voice reference.
Formal listening research often separates technical quality from suitability for a task. Your team does not need a laboratory to benefit from the same discipline. Keep playback conditions consistent, use the same script, and have reviewers judge independently before discussing the result.
Save the voice with its creative reference
Once approved, save more than the output file. Keep:
- the final structured brief
- the exact audition script
- the selected voice
- the rejected traits and decision note
- language and pronunciation decisions
- speed and production settings
- one approved reference export
- reviewer names and approval date
In Vois, Voice Design can turn selected attributes into a reusable voice, and the voice library keeps approved choices available to the production workflow. Use a stable name that describes the role, not the experiment number. "Museum guide, calm English" is more useful six months later than "Voice test 14."
Review the designed voice in every target language
A saved identity does not remove the need for language review. Rhythm, formality, pronunciation, and perceived age or warmth can shift across languages. Language tags also distinguish variants such as region and script, which matters when a project serves more than one locale.
For each target language:
- Translate the audition for meaning and natural spoken delivery, not word-for-word symmetry.
- Include names, numbers, and terms that are genuinely difficult in that locale.
- Ask a native or fluent reviewer to assess pronunciation, register, and character fit.
- Record approved pronunciation rules and any locale-specific exceptions.
- Keep the same role criteria, but allow the delivery to sound natural in the target language.
The pronunciation dictionary is useful for approved names and specialist terms. Do not use it to force one language's rhythm onto another.
Keep consent and impersonation outside the creative brief
Voice Design should create a voice from selected qualities, not covertly reproduce a real person. Avoid requests to imitate a public figure, colleague, customer, or performer without a clear right and reason to do so. If a project uses a cloned voice, keep documented consent and follow the voice cloning ethics guide.
The FCC has ruled that AI-generated voices in robocalls fall under restrictions governing artificial or prerecorded voices. That is one example of a broader point: a creative tool does not remove the legal and operational duties attached to distribution. Review the actual use, audience, channel, and jurisdiction.
A reusable Voice Design template
Copy this brief into the project before generating candidates:
Role: [What the voice does]
Listener: [Who hears it and in what context]
Delivery: [Relationship, energy, and emphasis]
Pace: [Speed, pauses, and sentence rhythm]
Texture: [Audible vocal qualities]
Emotional range: [Allowed variation]
Avoid: [Specific unwanted traits]
Audition must test: [Names, numbers, emotions, long sentences, target languages]
Approval test: [What reviewers must agree is true]
A concise, testable brief makes exploration faster because the team knows what to reject. The goal is not to generate the most impressive voice in isolation. It is to approve the voice that performs the intended job reliably.
Sources
- ITU-T P.808, Subjective evaluation of speech quality with a crowdsourcing approach
- RFC 5646, Tags for Identifying Languages
- NIST AI 600-1, Generative Artificial Intelligence Profile
- FCC, AI-generated voices in robocalls
A consistent custom voice begins with a precise decision, not endless generation. Explore Voice Design, save the approved reference, and get started when your team is ready to build a repeatable voice workflow.
The Vois Team