How to make a two-voice ad script, a customer and an agent, for a draft fast

This page may contain affiliate links from which Blur Studio earns commissions.

A two-voice ad, a customer stating a problem and an agent (or the product itself) answering it, is a reliable format because it mirrors how a real sale actually happens: objection, then resolution. The slow part has always been getting a draft version good enough to react to. Booking two voice actors for a script that might get rewritten twice before anyone approves it is expensive timing, not just expensive talent.

The usual cycle without this tool looks like: write a script, pitch it internally on the page alone, get a vague "sounds fine on paper" approval, book talent, record, and only then discover in the studio that a line everyone liked in writing doesn't land when it's actually said out loud. By that point you've spent the studio time and the actors' fee finding that out. Hearing a voiced draft before any of that happens moves the "does this actually work" moment to the cheapest possible point in the process, which is the entire value of doing this at the script stage rather than skipping straight to production.

What this tool is actually good for here

ElevenLabs Text to Dialogue generates a multi-speaker script as one audio file, with a consistent voice per speaker, built on the Eleven v3 model so the back-and-forth has real tonal shifts rather than two flat voice clips. For an ad draft specifically, that means you can test timing, tone and the actual words days before anyone commits to a studio booking.

Two limits to plan around: 2,000 characters per request (plenty for a 30- or 60-second ad script), and up to 10 distinct voices per request, though a two-person ad only needs two. The thing that actually matters for an ad specifically, more than for internal training audio: commercial use requires a paid plan. ElevenLabs' free tier doesn't include commercial license rights; that starts on the Starter plan and up, so a free-tier draft that works gets you excited about a script you then can't legally run as-is.

Text to Dialogue also reads bracketed delivery tags and punctuation as performance direction, not just words. A tag like [skeptical] or [reassuring] placed before a line, or an ellipsis to trail off, shapes the read more reliably than writing "say this warmly" as a stage direction, which the model will otherwise just read aloud. For a two-voice ad specifically, that's the difference between a customer line that sounds genuinely doubtful versus one that reads the doubt flatly, which matters more for an ad than for almost any other use of this tool, since the whole format depends on the objection sounding real before the resolution lands. The full tag list is in ElevenLabs audio tags and v3.

How to make the draft, step by step

1. Write short, punchy lines. Ad scripts read differently out loud than they look on a page; keep each line close to how someone would actually talk, not written-copy phrasing. If the read still sounds flat, see making an AI voice sound less flat in ad scripts.
2. Pick a voice for the customer and a different, complementary one for the agent, thinking about contrast (age, tone, pace) the same way you would for a two-actor casting decision. For the customer voice specifically, resist the urge to pick the most polished-sounding option in the library; a voice that sounds a little too smooth and confident undercuts the whole premise of someone genuinely hesitant, and a slightly more ordinary, relatable voice usually sells the objection better than a broadcast-quality one does.
3. Generate the audio through ElevenLabs' Text to Dialogue tool and check the length against your actual ad slot, 15, 30 or 60 seconds. AI-voiced dialogue often reads slightly differently in pacing than you'd expect from reading the script silently, and a script that looked like a tight 30 seconds on the page can easily generate at 34 or 35, which matters if you're buying a fixed ad slot that won't accept an over-length file. Pacing can be tuned as described in adding pauses and whispers to an AI voiceover.
4. Get sign-off on the script and performance, not just the words. A client approving a flat read of good copy isn't the same as them approving how it'll sound with real delivery. Send the actual audio file, not just a transcript of what you generated, and ask them to listen with headphones on once before signing off, since approving a script someone skimmed on a phone speaker in a meeting is a weaker approval than it looks.
5. Decide, before the final cut, whether this voice is the final voice. That's a brand and legal question as much as a quality one, covered below.

Two variants worth writing differently. A 15-second radio or social spot has room for essentially one objection and one resolution line each, so cut the script until the customer's line states the problem in under 10 words; anything longer and the ad feels rushed before the agent even responds. A 60-second video ad can afford a short back-and-forth, two or three exchanges, which gives you room to show a second, smaller objection getting handled too, often the more convincing structure for a considered-purchase product (real estate, enterprise software, higher-ticket services) where one objection and one answer feels too neat to be believable.

What I'd actually tell someone making an ad with this

For getting a script approved fast, this is genuinely one of the best uses of the tool: it turns "trust me, this will sound good" into "listen to this." The same idea works for pitches, as in making quick voice drafts for a pitch demo. Where I'd slow down is the jump from draft to final. An ad is public, repeated, and attached to your brand in a way an internal training clip isn't. Before you ship an AI-voiced ad as final, two things are worth a straight answer, not an assumption: does your market or platform have any disclosure expectation around AI-generated voices in advertising, and does the specific voice you used carry any licensing restriction beyond the ElevenLabs plan you're on. Neither is usually a dealbreaker, but "I didn't think to check" is a worse position than deciding on purpose.

There's a specific version of this worth naming for a real estate or property ad: if the "customer" voice in your script sounds like it's expressing a real buyer's genuine hesitation and the ad implies it's an actual client testimonial rather than a scripted dramatization, that's a different and more serious problem than a disclosure question, since it risks presenting a fabricated voice as a real person's endorsement. Keep the framing honestly dramatized, a scripted scenario, not a disguised fake testimonial, and this entire concern goes away.

When to skip this

If the ad is genuinely small-stakes, a local business's social clip, not a paid campaign, shipping the AI-voiced version as final is reasonable. If it's a paid campaign running at any real spend, the cost difference between the AI draft and a real voice actor for the final cut is usually small relative to the media spend around it, and the downside of it sounding or being flagged as synthetic in the wrong context is the bigger risk to manage, not the production cost.

There's also a scale threshold worth watching for: if you're producing ad variants regularly, testing five versions of the same script against different audiences, swapping in a different product name or offer each time, doing this by hand in the dashboard becomes the bottleneck, not the voicing itself. That's the point where scripting the generation through the API, rather than the web interface, starts paying for the setup time within the first few variants, especially once you're also tracking which variant actually performed best and want that data feeding back into the next round of scripts automatically rather than someone compiling it by hand.

How I'd approach this

ElevenLabs' plans are the right tool for exactly what this page covers: a fast, cheap, approvable draft. If you're running ad production regularly and want a repeatable pipeline, script variants generated and tested automatically, voice selection standardized across a brand's whole ad library, that's a build I do for clients who've outgrown one-off drafting and want the whole loop to run without someone manually chunking scripts each time.

A version of that build I'd actually recommend to an agency running this for several clients: a shared voice library mapped to each client's brand (so nobody has to remember which voice belongs to which account), a script-approval step before anything generates, and automatic variant generation for A/B testing a few different objection lines against the same resolution. That turns a manual, per-request workflow into something a small team can run across a dozen accounts without the process itself becoming the bottleneck.