A robotic voice is a voice with no hesitation in it.

The word describes prosody, not quality. A synthetic voice can be warm, clear and pleasant and still sound robotic, because the pattern of stress and pause does not match how a person speaks.

Human speech is uneven. We speed up through familiar phrases, slow down on numbers, take breaths mid-thought, and change pitch within a sentence to signal what matters.

Synthetic speech is typically even: the same care applied to every syllable, the same pause length at every comma, the same fall at every full stop.

That evenness is what the ear reads as machine, and it survives improvements in timbre. A high-quality voice model with flat prosody still sounds synthetic.

These are the features worth listening for, in the order they are usually noticed.

Even a good voice model mispronounces the things that matter most to a business: names, products, places and numbers.

Heteronyms are the predictable failure: words spelled the same and pronounced differently, such as read, lead, live and tear. A model picks the more common pronunciation, which is wrong half the time.

Acronyms split into spelling or read as words depending on context, and the model decides without knowing your organisation's convention.

Numbers have their own rules: years, prices, ranges, phone numbers and decimals each have a spoken form. Check every one in your script, because these are the errors a listener remembers.

Write a test script that contains your real content: your product name, a price, a date range, an acronym, a sentence with a semicolon, and a paragraph longer than three sentences.

Generate it, then listen on phone speakers and on headphones. Most listeners use the former, and it exposes pacing problems that headphones hide.

Have someone else listen without the text in front of them and write down what they heard. Mishearings are the useful data, and they are frequent with names and numbers.

Fix the script before you fix the voice. Adding a comma, spelling a name phonetically or splitting a sentence solves more than switching engines.

Vary the sentence length in the script. A voice model has more to work with when the input is not eight sentences of identical structure.

Break the script into short segments and set pauses deliberately, rather than relying on punctuation to generate them.

Use the pronunciation dictionary for names, acronyms and heteronyms. This is a five-minute job that removes most of the embarrassing errors.

Add slight variation between takes by regenerating a sentence and choosing the better one, rather than accepting the first pass across a whole script.

Apologies, condolences, safety instructions and anything with a legal consequence are better in a human voice. Listeners forgive synthetic narration of facts and do not forgive synthetic empathy.

Anything that must be trusted as a person speaking to them: a message from a named individual, a personal response, a customer service call. Impersonation is the problem, not the quality.

Cloning a specific person's voice requires their informed consent, in writing, and a scope. Without that it is a deception regardless of how good it sounds.

If the content is a disclosure, an accessibility feature or a translation where no human voice is available, synthetic is reasonable and often better than nothing, provided it is clear what it is.

Where a listener could reasonably believe they are hearing a person, say that the voice is synthetic. Platforms increasingly require it, and audiences increasingly expect it.

A short line at the start or in the description is enough, and it costs less than the alternative: being caught.

Do not use a synthetic voice to imply a human is present when the call or message is automated. The voice is not the problem there; the claim is.

Keep a record of the engine, the settings and the consent for any cloned voice. It answers the question that eventually arrives.

Long-form narration of factual material: documentation, courses, product walkthroughs. The listener wants information and the pacing can be tuned.

Short system prompts and accessibility reads, where consistency and clarity matter more than warmth.

Localisation, where a synthetic voice in a language you cannot staff beats English-only audio for that audience.

Bulk variation, such as hundreds of short announcements, where a human recording would be impractical and the content is homogeneous.

The common professional setup is synthetic for the body and human for the moments that carry emotion: an introduction, a closing, an apology.

Match the room. Record the human segments in a space that matches the synthetic ones, or treat both with the same reverb, so the switch is not audible as a change of place.

Keep the same voice for a series. Listeners build familiarity quickly and a change mid-series reads as a mistake.

Label the two clearly in the edit, because a switch is often the moment a listener decides whether to trust the whole recording.

Keep the human segments short. Thirty seconds of a real voice at the start of a nine-minute narration sets the listener's expectation, and the synthetic body is then judged as narration rather than as an imitation of the person they just heard.

Speech rate is the setting listeners notice most and adjust most. Most engines default slightly fast for long-form listening, so test at a slower rate for anything over five minutes.

If your player allows speed control, keep it. Listeners who use screen readers or who process audio slowly rely on it, and locking playback speed turns an accessible asset into an unusable one.

Pair the audio with a transcript. It serves deaf and hard-of-hearing readers, it makes the content searchable, and it gives you a place to fix a name the voice mangled without re-recording anything.

Check that any captions are timed to the synthetic pacing rather than a human read. Generated speech often has shorter gaps than a transcript file assumes, and captions drift within the first minute.

Play one minute on phone speakers with the script in front of you, and mark every word that surprised you.

Then play it to someone else and ask what they heard, without showing them the text. Mishearings collect in names, numbers and acronyms.

Then listen for breath alone. Its absence or misplacement is the single cheapest improvement available, and it is a script-level fix.

Use the review list on this page to run the same checks across a batch, so the second recording is as good as the first.