Voice design vs voice cloning: when to pick which

Voice design and voice cloning both produce a re-usable Voice row in your Hakim library. Both feed the voice field on every subsequent POST /v1/audio/speech call. Both work across the three TTS tiers (Saree' / Faseeh / Mulhim) and every supported language. From a hundred feet up, they look like two paths to the same end.

They are not. The two surfaces are designed for completely different jobs, carry different consent + legal postures, and the failure mode of picking the wrong one shows up either as wasted budget or, occasionally, as a regulatory problem. This post walks through the decision rules, the consent posture for each path, and the use-case patterns we see hold up in production.

If you only have a minute: voice cloning is for "this specific person's voice"; voice design is for "a voice that matches this description". If your answer to the brief is a name, you want cloning. If your answer is a paragraph, you want design.

The two paths at a glance

The rest of this post walks through the dimensions in detail.

When voice cloning wins

The brief: there's a specific real human whose voice is the asset you need to replicate. Five product surfaces where cloning is almost always the right answer:

• Founder + executive voices. A SaaS company that wants their CEO's voice on the welcome email, the keynote intro, and the in-product onboarding tour. The CEO records ten seconds once; the cloned voice carries every subsequent piece of content. The person is the brand.
• Hired narrators on a long-running show. A podcast network with a roster of hired hosts wants to render new episode intros without scheduling each host into a studio. The host records ten seconds with full consent; the clone produces season after season.
• Audiobook narration with a series voice. A publisher producing a 12-book trilogy needs the same narrator voice across all twelve volumes, including any future installments. Cloning guarantees the timbre stays identical across years of production without re-booking the narrator.
• Employee voices for internal product. A field-tech app where technicians get audio walkthroughs in their own voice (more trust, more clarity than a generic voice). Each technician clones their own voice on signup.
• Dubbing actors at scale. A localisation studio dubbing a TV show into Arabic across MSA + four dialects. The original actor signs the rights; the clone produces every dialect's version without re-casting per language.

The thread connecting these: the value the team is buying is that specific person's voice as part of their identity. The voice isn't swappable; it's the asset. Voice cloning is the only surface that preserves that.

# Voice cloning request shape
curl -X POST https://api.tryhakim.ai/v1/audio/voices \
  -H "Authorization: Bearer $HAKIM_API_KEY" \
  -F "name=Aya, narrator" \
  -F "language=ar" \
  -F "consent_confirmed=true" \
  -F "file=@./aya-10s.wav"

Note consent_confirmed=true. That flag is non-optional on the cloning endpoint and binds your account to the legal attestation that you have documented consent from the person whose voice you uploaded. If the sample is later flagged (because, for example, it's reported as a public figure's voice without authorisation), the audit trail will reach you.

When voice design wins

The brief: you need a consistent, brand-aligned voice but there's no specific real human whose voice you're after. Five product surfaces where design is the right answer:

• Brand voices for early-stage products. A team that doesn't have a hired voice actor on retainer but needs a consistent brand voice across audio notifications, in-product help, and marketing reels. Design five candidates on a Tuesday afternoon, pick the keeper.
• Character voices for games and interactive media. A studio shipping a game with twenty named characters can't realistically book twenty voice actors. Design produces twenty distinct, consistent timbres that the studio owns outright.
• Localised brand voices across dialects. Need the same brand personality in MSA, Khaleeji, Egyptian, and Levantine? Write the same prompt with a dialect cue per language. The brand timbre stays consistent; the dialect shifts. Cloning would require recording the same actor in four dialects · expensive and rarely produces natural results.
• A/B testing voice fit. Marketing teams testing whether audiences respond better to a warm narrator vs an authoritative one can ship both variants at near-zero production cost. Cloning would require recording two different humans · prohibitive.
• Gap-fill in the preset catalogue. Need a "young Bedouin storyteller with a desert wind in their voice" and we don't ship a preset that matches? Design it. Cloning would require finding that exact person, which is the kind of casting problem voice AI was built to avoid.

The thread connecting these: the value the team is buying is a voice that matches a description, not a specific person's voice. The voice is fungible up to its capability profile (timbre, register, pacing). Voice design is the only surface that produces that without the casting overhead.

# Voice design request shape
curl -X POST https://api.tryhakim.ai/v1/audio/voices/design \
  -H "Authorization: Bearer $HAKIM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "description": "warm raspy elderly woman, low-pitched, magnetic, slight breathy tremor",
    "sample_text": "Welcome. This is a quick sample of the voice you described.",
    "save_as_voice": { "name": "Storyteller Demo" }
  }'

Note there's no consent_confirmed field. There isn't a real person to consent to anything. The legal posture for the resulting voice is "this is a synthetic timbre produced by Hakim Mulhim v3 from a text description; ownership terms are governed by the standard Hakim Voice Library terms" · which means you own it as part of your account, no third-party claims attach.

Consent + legal posture

This is the dimension where the wrong choice has the highest blast radius. Picking voice cloning when you should have used voice design costs you legal-review time. Picking voice design when you should have used cloning costs you authenticity. Neither is a project-ending mistake, but the cloning-when-you-shouldn't side is the higher-risk direction.

Voice cloning

Every cloning call requires consent_confirmed=true and records:

• The org id and user id of the API key holder.
• The IP address and timestamp of the cloning call.
• The MD5 of the audio sample (so re-uploads of the same sample are detectable in the audit log).
• The voice id assigned at training time.

Public-figure samples are auto-flagged. Our moderation pipeline runs the sample against a registry of known voices; matches above a similarity threshold get held in pending_manual_review until a human at Hakim either confirms the sample (e.g. a journalist uploading the politician they have a documented interview consent with) or rejects it. The audit trail reaches the originating org either way.

This posture is consistent with the EU AI Act's high-risk biometric category and provides the documented chain-of-custody that downstream regulators expect. For production deployments in regulated industries (financial services, healthcare, government communications) the auditable consent posture is the cloning surface's primary differentiator over ad-hoc clone tools.

Voice design

Voice design's consent posture is simpler because there is no third party. The prompt is text you wrote; the resulting voice is a synthetic timbre Mulhim produced. The legal stack:

• The prompt is logged for audit (you can revisit "what did I ask for?" months later, useful for brand-consistency reviews).
• The resulting voice belongs to your org account · standard Hakim Voice Library terms apply.
• Public-figure prompts are still flagged for manual review, because "a voice that sounds like Morgan Freeman" is a moderation vector even when no audio source is uploaded.
• There is no third-party consent to gather, document, or revoke · because there is no third party.

The simpler legal stack is one reason voice design ships on a premium tier; the underlying infrastructure is more expensive, but the integration cost on the customer side is also lower. No consent forms, no actor agreements, no per-territory rights questions. The voice exists because you described it; it didn't exist before.

Cost + latency differences

For a team comparing cost between the two paths, the calculus comes down to:

The "marginal voice cost" line is the one that surprises teams new to voice design: a saved voice is the result of 3–5 design iterations (the iteration loop is the unit of design; one-shot designs rarely land). Budget accordingly. Once saved, though, the voice renders speech at the same per-character cost as any other voice · the design premium is paid up front, not on every render.

Latency posture differs too:

• Cloning produces a voice asynchronously · the POST /v1/audio/voices call returns immediately with a processing status. The clone trains in the background; voice_clone.ready webhook fires when the voice is usable (typically a few seconds for a 10-second sample).
• Design is synchronous · the design endpoint returns the rendered audio sample inline in the response. The voice is usable the moment the response lands.

For teams shipping a self-serve voice surface (e.g. "let our users create their own voice in our app"), the synchronous behaviour of design produces a better UX than the async-with-webhook shape of cloning. For backend pipelines where the voice gets used hours or days later, either works.

How they interoperate

The two paths aren't mutually exclusive in a product. We see teams running both in production for different use cases. Three patterns worth knowing:

Pattern 1 · Cloned founder voice + designed brand voices

The CEO clones their voice once; that voice carries personal content (Loom-style product updates, town-hall recaps, internal all-hands intros). The marketing team designs five brand voices for different audio surfaces (TikTok ads, podcast preroll, in-product notifications, email-to-audio digests, IVR). The audit trail keeps the personal voice clearly separated from the synthetic brand voices; the cloned voice is reserved for personal-attribution contexts.

Pattern 2 · Designed prototype → cloned hero

A brand isn't sure what their voice should be yet. The team designs ten candidates, ships an A/B test, and learns the warm female podcast register lands hardest with their audience. Six months later, they hire an actual voice actor matching that register, clone the actor's voice with consent, and migrate production traffic to the cloned voice. The designed voices stay in the library for legacy content; new content uses the human.

Pattern 3 · Cloning at the centre, design at the edges

A dubbing studio clones the original-language actors with full consent (the hero work). For minor characters or crowd voices where casting individual actors is impractical, they design voices instead. The audit log keeps the two cohorts clearly labelled (kind: cloned vs kind: designed); per-character rights deliveries reference only the cloned voices.

Per-use-case decision tree

Frequently asked questions

Can I use voice design to imitate a real person? Prompts that reference real people trigger our moderation pipeline. "A voice that sounds like Morgan Freeman" gets flagged for manual review, and even when approved the result is influenced by the description's general texture rather than the specific person. To get a specific real person's voice, use voice cloning with proper consent. Imitation is not a feature; it's an anti-pattern.

Can I clone my own voice and then "edit" it via a voice prompt? Today the two paths are distinct · a cloned voice is the cloned voice. There's no "designed-over-cloned" composition surface. You can layer non-verbal tags ([breath], [laugh]) on top of any saved voice via the speech endpoint, but a voice_prompt field on Mulhim doesn't reshape a cloned voice's timbre · it adds a tonal overlay on top. For full timbre reshaping, design a new voice.

How do I decide which path to use for our brand voice? If you've already booked a voice actor and have their consent, clone them. The booking is the expensive step; the consent is sunk cost. If you haven't booked an actor yet, design candidates and ship them on a real product surface before committing to a hire. Most teams that design first find out their assumed-best voice isn't actually what audiences respond to; you save the actor budget for the right person.

Can I save designed voices and cloned voices in the same library? Yes. Same Voice table, addressable by the same voice field on the speech endpoint. The only public difference is Voice.kind ("cloned" vs "designed") which surfaces on the voice picker so operators can tell them apart. Both kinds inherit the same dialect support, sample-rate options, and per-tier behaviour.

Does cloning produce better audio than design? On average, yes · cloning has a real reference audio signal to optimise against, which is more information than a text prompt can carry. The gap closes when the prompt is well-engineered and the iteration loop is run properly (see the prompt engineering playbook). For production audio where the "is this a real person?" test matters, cloning still wins. For brand voices, character voices, and fictional speakers, the gap is irrelevant.

Can I share saved voices across orgs? No. A Voice row is scoped to the org that created it. For cloning, this matches the consent scope (the person consented to your org, not to other orgs cloning the same voice). For design, this matches the audit trail (the prompt belongs to your org). If your business requires multi-org access (e.g. an agency creating voices on behalf of clients), use the Hakim Enterprise plan which supports per-client voice scoping.

What this commits us to

Voice design and voice cloning are both first-class surfaces. Neither is sunsetting; neither is the "right" answer to every voice problem. The teams that ship voice products that age well are the ones that treat them as complementary tools with different consent postures and different cost shapes.

We invest in both. Voice cloning has been a Hakim capability since v1.0; voice design ships on the new Mulhim tier and gets the same production discipline as the rest of the stack · audit logging, moderation pipeline, multi-region inference, the works. The voice-cloning runbook for the EU AI Act documents the regulatory posture for cloning; the voice prompting explained post documents the same for design.

If you're starting from scratch and don't have a specific person in mind, open the playground on Mulhim v3, the worked example loads automatically, and you can hear what voice design produces in 30 seconds. If you have an actor or speaker in mind, head to the voice cloning endpoint with their consent form ready.

Voice is the next interface. The team that ships it well picks the right surface for each voice, not the same surface for every voice.