How to Choose an AI Voice for Ads, IVR and Explainer Videos: Checklist and Comparison Table

Short answer: Pick the voice by ear, using your real script, on the device your audience will use. Generate the same 50-word script with three or four voices, add a no-style control, change one thing at a time, and score what you hear against a short list of criteria. The right voice for an advert is rarely the right voice for a phone menu.

This article gives you that process, a comparison table for the three most common jobs, and a scorecard you can copy. The examples come from the demo clips on voiceover.my, where every clip shows the transcript, the voice and the request behind it.

Start with the job, not the voice

Before you listen to anything, write down three things.

1. Where it will be heard. A phone line, a phone speaker, a laptop, a shop speaker, a car.
2. What the listener must do next. Press a key, visit a place, understand a process.
3. How long it must run. Fifteen seconds, a minute, five minutes.

These decide more than the voice does. A friendly, fast read is good for a social advert and bad for a menu with five options.

What each job needs

Qualitative guidance by voiceover.my, 16 October 2026. It reflects common production practice and our own tests, not a published standard.

How pace varies in real clips

We measured the pace of six reference clips from our Malay gallery. The scripts are 50 to 58 words long, written for six situations, and generated with the settings shown on each demo page.

Calculated by us from clip durations and word counts in our gallery data, 16 October 2026. Pace depends on the script, voice and style, so these are a reference for planning, not a rule. The Malay scripts were written for the situation, not for timing.

One thing stands out. The podcast-style clip ran about 50 percent faster than the news clip, which fits a conversational read. The solo clips sit in a narrower band of roughly 107 to 127 words per minute. Use these as a rough planning baseline, then time your own script.

You can open each of these on the demo pages to see the exact transcript, style line and request:

The scorecard

Print this, or copy it into a spreadsheet. Score each voice from 1 to 5 after listening to the same script. Add a note for each score.

A voice that scores 5 on tone and 2 on pronunciation is a rewrite or a different voice, not a pass. Decide your must-haves before you start (for most Malay projects, accent fit and pronunciation) and reject anything that fails them.

A repeatable test method

This is the method we use to compare voices. It follows one rule: change one variable at a time.

1. Pick one script of about 40 to 60 words that looks like your real one, including your hardest names and numbers.
2. Generate a control with each voice and no style instruction.
3. Choose the best two by ear, not by name.
4. Add one style line such as calm and clear, and compare against the control.
5. Change the model (Flash-Lite versus Flash) only after you have picked a voice, so you know what changed.
6. Listen on the target device: a phone speaker, a laptop, earbuds.
7. Record the decision: model, voice, text, style and your score.

Because every voiceover.my demo shows the request used, you can copy the starting point for any of them. Here is the shape of a request for a phone-menu prompt, using the style line and voice from our customer-service demo. The text is shortened from the full demo script.

{
  "model": "gemini-3.8-flash-lite-tts",
  "input": [{
    "type": "user_input",
    "content": [{
      "type": "text",
      "text": "Terima kasih kerana menghubungi kami. Untuk menyemak tempahan, tekan satu. Untuk bantuan pembayaran, tekan dua.",
      "annotations": [{ "type": "speech_metadata", "style": "clear, polite, professional" }]
    }]
  }],
  "response_format": { "type": "audio" },
  "generation_config": { "speech_config": [{ "voice": "Sulafat" }] }
}

This follows the request structure in Google’s documentation. Delete the annotations entry to produce your control.

Choosing for adverts

• Hook first. The first three seconds decide whether the listener stays. Generate two or three versions of your opening line alone.
• Warmth before polish. A slightly imperfect friendly read often works better than a flawless cold one.
• Keep the call to action short and clear. Test it on a phone speaker, because many people will hear it that way.
• Use the same voice across a campaign unless variety is the idea.

Our reference advert is a short café script with the style line upbeat, clear, friendly, generated with the Leda voice. The reviewer recorded that it fitted the use and sounded Malaysian with clear Malay. That is one person’s notes on one clip, so listen to it yourself before you reuse the style.

Choosing for IVR

• Clarity beats personality. The caller is trying to do something, not enjoy a voice.
• Keep each prompt short. Our sample menu uses short options and ends with “tekan sifar” to repeat.
• Match your phone system’s format. Google documents mu-law and A-law output as well as raw PCM, selected through the response format. WAV is the default for single requests. Confirm which one your system needs.
• Test on a real call. Telephone compression changes how voices sound.
• Plan for edits. Prompts change. A text-based workflow makes this cheap, which is where AI voices help most.

Choosing for explainer videos

• Match the pace to the visuals. Time the script before you edit the video.
• Avoid monotone over a long run. Listen to the whole piece. A voice that is pleasant for 20 seconds can tire over three minutes.
• Use short sentences and clear signposts such as “first”, “next” and “finally”.
• Add pauses deliberately. Google documents inline tags such as <short pause> for short pauses.

Our explainer sample is a short booking walk-through with the style line friendly, clear, steady pace, generated with the Despina voice. Treat it as a starting point to compare against a no-style control, not as a proven recipe.

Flash-Lite or Flash?

Last verified 16 October 2026. Sources: Flash-Lite TTS page, Flash TTS page, pricing. The last row is our own working choice for tests, not a Google recommendation for Malay.

Both use the same request format, so you can switch with one parameter and compare on your script.

Common mistakes

• Choosing from a one-line preview. Preview with your real script.
• Changing voice, style and model at once. You will not know what helped.
• Skipping the control. Without a no-style version you cannot tell whether the style did anything.
• Judging only on headphones. Your audience may use a phone speaker or a call line.
• Trusting a language label as proof of accent. Standard Malay is supported by the models, but a local-sounding result needs a local listener.
• Forgetting the licence. Check the service terms for commercial use before publishing.

The 10-point checklist

• I know where the audio will be heard.
• I have one realistic 40 to 60 word script.
• I generated 3 to 4 voices without style.
• I picked the best two by ear.
• I added one style line and compared to the control.
• A Malaysian listener heard it blind.
• I tested on the target device.
• I checked names, numbers and loanwords.
• I checked the audio format my platform needs.
• I recorded the model, voice, text, style and decision.

Ready to hear examples by use case? Start with the customer-service demos, then compare with advertising and explainers.

For the background on how Malay voices work and what to expect from accent and cost, read our complete guide to Malay text-to-speech. If you are weighing AI against a person, see AI voice-over versus a human voice actor in Malaysia.

Sources

All sources below were accessed on 16 October 2026.

• Speech generation (TTS), Gemini API documentation, for request format, output formats, inline tags and limitations.
• Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS, for model positioning.
• Gemini Developer API pricing, for audio output rates.