How to make a sample customer call recording for training, without hiring actors

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

New hires in any phone-heavy role learn call handling one of two ways: they listen in on real calls for a few weeks and pick it up by osmosis, or someone hands them a script and hopes. Neither is fast. A short, scripted example of a call going well, and one going badly, gives a new hire something concrete to compare their own calls against on day one, not week three.

The obstacle is usually production, not writing. Getting two people in a room to record a believable back-and-forth, or paying a couple of voice actors for a two-minute clip, is disproportionate effort for something you'll use in an onboarding deck and then mostly forget about.

What ElevenLabs Text to Dialogue actually does

Text to Dialogue is ElevenLabs' feature for turning a multi-speaker script into one piece of audio, with each speaker's lines voiced separately and strung together automatically. It's built on the Eleven v3 model, which is what gives it the ability to handle emotional inflection and natural turn-taking rather than sounding like two separate text-to-speech clips glued together.

Two limits matter in practice. First, there's a 2,000-character limit per request, counting every speaker's lines combined, so a two-minute call script will usually fit in one request but a longer one won't. Second, while there's no cap on how many lines or turns a dialogue can have, a single request can use at most 10 distinct voices. For a two-person customer call, neither limit is close to a problem.

Two features worth knowing about before you start writing. First, Text to Dialogue reads punctuation and bracketed audio tags as delivery instructions, not just words to read aloud: an ellipsis makes a line trail off, a tag like [sighs] or [frustrated] before a line changes how it's performed, and this is a more reliable way to direct tone than trying to describe the mood in a separate note. Second, there's an audio-effects setting that places the whole conversation in a scene, such as a phone call or a call-center room, which adds real ambience rather than a flat, dry studio read. For a customer-call training clip specifically, turning this on is worth the extra minute, since a call that sounds like it's actually on a phone line is more believable to a new hire than two clean, close-mic'd voices.

How to make the recording, step by step

1. Write two short scripts, not one. A "good call" version where the agent handles the issue calmly and correctly, and a "bad call" version with the mistakes you actually see your team make (interrupting, guessing instead of checking, no acknowledgement of frustration). Contrast teaches faster than a single good example.
2. Mark every line with who's speaking. Text to Dialogue needs each turn attributed to a speaker before it will assign voices correctly. The same labelling approach is covered in voicing a script with several characters.
3. Pick one voice for the customer and one for the agent, and keep them consistent across both the good and bad versions, so the comparison is about the conversation, not about different-sounding voices confusing the listener.
4. Generate the audio through ElevenLabs' Text to Dialogue tool and listen the whole way through once before sharing it. Scripted dialogue occasionally produces an odd pause or a line read with the wrong emphasis, and catching that before training day is easier than explaining it away live.
5. Use it as a discussion starter, not a standalone lesson. Play the bad version, ask the room what went wrong before you tell them, then play the good version as the comparison.

Two practical notes that save a redo. 2,000 characters works out to roughly 300-350 words of dialogue, which is a natural two-minute, two-person call. If your scenario genuinely needs to run longer, for example a multi-step troubleshooting call, generate it in two linked clips rather than trying to cram it into one request: keep the same two voices and settings in both clips, and the join in your audio editor is seamless, since it's the same two voices picking up where they left off, not a continuity problem a listener will notice.

Voice selection matters more than any setting you'll tweak. Pick two voices that are easy to tell apart at a glance, not just two voices that both sound fine on their own: a visible age or tone contrast works better than two similar mid-range voices, since a listener losing track of who's speaking mid-sentence is the single most common way training audio fails, and it happens exactly when half the room has already stopped paying close attention.

A second scenario: the escalation call

The good-call/bad-call pair above works well for a routine interaction. For a field service or dispatch team, the scenario that actually needs practice is usually angrier: a customer calling about a missed appointment window. Here's how I'd script that one differently. In the bad version, the agent apologizes once, then immediately moves to rescheduling without acknowledging why the customer is frustrated, which reads as dismissive even though the agent technically "handled" the call. In the good version, the agent names the frustration directly ("I understand you took time off work for this, and that's on us"), explains what happened in one sentence without over-explaining, then offers a specific, concrete next step (a time window, a discount, a priority slot) rather than a vague "we'll make it right." The script difference between those two versions is small on paper, maybe 15 words, but it's the exact gap that separates a call that de-escalates from one that doesn't, and hearing both back to back makes that gap obvious in a way that a written script alone doesn't.

What I'd actually tell a training manager trying this

This is a genuinely good fit for exactly what it's being asked to do here: a short, repeatable example clip for onboarding decks, refreshed occasionally as your product or policy changes. Where I'd push back is if someone starts treating it as a training program rather than a training aid. A single scripted good/bad pair teaches the shape of a good call. It does not teach someone to think on their feet when a real customer goes off-script, gets angry in a way your example didn't anticipate, or asks something you didn't write a line for. If your onboarding currently has zero example audio, this closes that gap well. If your onboarding has zero live call shadowing or live practice, that's the bigger gap, and no amount of AI-generated dialogue fixes it. For written material, a staff manual can become short audio lessons.

One more thing worth building into the habit rather than treating this as a one-off: refresh the clip when your policy or product actually changes, not on a fixed schedule. A good-call example that still references a return policy you've since dropped, or a product that's been discontinued, teaches a confidently wrong habit, and because the audio sounds polished and authoritative, new hires tend to trust it more than a written policy doc that's visibly out of date. Since regenerating it costs you a few minutes rather than rebooking two actors, there's little reason not to keep it current. The same customer and agent setup also works for an ad with two voices.

When to skip this

If your team is small enough that two colleagues can record a rough example on a phone in five minutes, that's genuinely fine too, and probably faster than writing a script tight enough for Text to Dialogue's character limit. Where this tool earns its place is when you need several variations (different scenarios, different products, different languages) and don't want to re-book two people's time for each one.

One case where I'd slow down before using this at all: if your industry has required disclosure language, a collections call, certain healthcare or financial scripts, a regulated-industry greeting, get whoever handles compliance to sign off on the "good" script before you record it. A believable, well-performed call that quietly skips a required disclosure teaches a new hire the wrong habit as if it were the right one, which does more damage than having no example audio at all. This isn't a limit of the tool, it's a limit of treating any scripted example, generated or recorded, as pre-approved just because it sounds right.

How I'd approach this

I build training and onboarding tools for clients fairly often, and audio like this is usually one small piece of something bigger: a structured onboarding path with quizzes, tracked completion, and real call data instead of scripted examples once the team is live. If you just need the occasional training clip, ElevenLabs' plans are the right tool and there's no reason to build anything custom. If you're trying to stitch this into a proper onboarding system, pull in real call recordings for coaching, or build it into a learning path your team actually tracks progress through, that's a custom build, and it's the kind of thing I do for clients who've outgrown a folder of training clips.