The Science of Voice in Learning: What Research Says About Retention

Research on voice in learning points to 4 findings. Spoken narration beats on-screen text when it explains a visual, but repeating the narration as on-screen text can hurt. Conversational wording helps learners apply ideas, and modern text-to-speech voices performed about as well as recorded human voices in recent studies.

Last updated: October 6, 2026. We read the published abstract of every study cited here on that date.

This guide is for course creators, instructional designers and L&D teams who want evidence, not opinions, about narration. Each finding below links to the original study. We also say where the evidence is thin, because several popular claims about voice and memory have no study behind them.

Key Takeaways

• Narration plus visuals beats text plus visuals in several experiments, including Mayer and Moreno (1998), where narration improved retention, matching and transfer scores.
• Don't put the full narration on screen next to a diagram or animation. Mayer, Heiser and Lonn (2001) found duplicated or summarized on-screen text lowered retention and transfer.
• Text plus narration can help when there is no other visual, as Moreno and Mayer (2002) found.
• Conversational wording improved transfer, not retention, in Mayer and colleagues (2004). Several voice effects show up on transfer tests more than on recall.
• Older synthetic voices hurt learning; newer ones didn't in the studies we checked. Craig and Schroeder (2019) found a modern text-to-speech engine performed about the same as a human voice in most respects.

On this page: Summary of studies · Narration vs text · Redundancy · Conversational style · AI vs human · Pacing · Limits · What to do · FAQ

Research on voice in learning

Most of the evidence comes from multimedia learning research, much of it by Richard Mayer and colleagues. Many of these studies used short narrated animations, such as an explanation of how lightning forms, followed by tests.

"Retention" in these studies means recalling what was presented. "Transfer" means using the idea to solve a new problem. Keep that difference in mind as you read the findings.

Narration vs on-screen text

When words explain a picture, diagram or animation, spoken words usually work better than the same words printed on screen. This is called the modality effect.

Mayer and Moreno (1998) showed students an animation of lightning formation or a car's braking system. Students who heard narration beat students who read the same words as on-screen text on retention, matching and transfer tests. The authors explain this with a dual-channel model: eyes and ears feed partly separate parts of working memory.

Mousavi, Low and Sweller (1995) found the same pattern with geometry worked examples. Presenting statements as audio, rather than as text next to the diagram, reduced split attention and helped learning across 6 experiments.

Reading while listening

It depends on what else is on screen. This is called the redundancy effect.

• With visuals, redundant text hurt. Mayer, Heiser and Lonn (2001) added on-screen text that summarized or duplicated the narration of an animation. Students did worse on retention and transfer than students who only heard the narration.
• Without visuals, it helped. Moreno and Mayer (2002) found students understood better with both spoken and written words, provided no other visual material was on screen at the same time.

So the rule isn't "never show text". Don't make learners read a paragraph while they are also trying to watch a diagram. Short labels on the diagram are a different case from full sentences.

The earlier version of this post called this the "split-attention effect". That was wrong: split attention is about having to look at 2 separate visual sources, while redundancy is about the same words arriving twice.

Conversational voice

Yes, for applying knowledge. Mayer, Fennell, Farmer and Campbell (2004) changed a narration from formal to conversational style, mainly by swapping "the" for "your" in 12 places.

Across 3 experiments, students who heard the conversational version scored higher on transfer tests. Their retention scores were not significantly different. In practice, write narration the way you'd explain it to one person: "your lungs", not "the lungs".

Show me: formal vs conversational narration

Formal: "The lungs take in oxygen and release carbon dioxide."

Conversational: "Your lungs take in oxygen and release carbon dioxide."

This is the kind of change Mayer and colleagues (2004) tested: mostly swapping "the" for "your". It is our own example, not a line from the study.

Human vs AI voices

Older research said yes; newer research says much less.

• Older synthetic voices did worse. Mayer, Sobko and Mautone (2003) found students did better on transfer, and rated the speaker more positively, with a human voice than with a machine-synthesized voice. Retention results were mixed.
• Accent familiarity mattered. In the same paper, a standard accent outperformed a foreign accent for those learners. Choose an accent your audience hears every day.
• Modern text-to-speech closed most of the gap. Craig and Schroeder (2019) compared an older engine, a modern engine and a recorded human voice. In most respects, learners using the modern engine didn't differ statistically from those hearing the human voice, in perceptions, learning or efficiency.
• Quality affected trust more than learning. Chiou, Schroeder and Craig (2020) found voice quality did not significantly change learning, but it did change trust and how learners saw the tutor.

The practical takeaway: a clear, natural-sounding voice matters, and learners notice poor quality even when test scores don't move. We found no study showing that synthetic voices make the brain treat content as "low priority", a claim the earlier version of this post made.

Learner-controlled pace

Yes. Mayer and Chandler (2001) let some learners control the pace of a narrated animation.

Learners with control did better on transfer tests, though not on retention. In your course, that means short segments with a clear "next" button and a player that can pause and replay.

A related finding from Mayer, Heiser and Lonn (2001): adding interesting but irrelevant details or video clips lowered transfer. Cut stories, sound effects and asides that don't support the learning goal.

Retention vs transfer

Be careful with headlines that say voice "boosts retention". Several of the findings above, including conversational style, voice type and pacing, showed up on transfer tests more clearly than on recall.

These studies also have limits:

• Many used short lessons in controlled settings. The lightning explanation in Mayer, Heiser and Lonn (2001) ran 140 seconds.
• Several used the same topic, how lightning forms, so results may not cover every subject.
• They were experiments, not full courses, so effects across a 10-hour program may differ.

We could not find primary studies for some popular claims, such as an ideal segment length in minutes or the effect of a specific pause length. We removed those claims from this post.

Applying research to courses

1. Narrate visuals instead of printing paragraphs next to diagrams and animations.
2. Keep on-screen text short when a visual is present. Use labels, not full sentences.
3. Offer full text when there's no visual, such as a reading with an audio version.
4. Write conversationally. Use "you" and "your", short sentences and everyday words.
5. Pick a clear voice with a familiar accent for your learners, and keep it consistent across modules.
6. Split lessons into segments and let learners control the pace.
7. Cut extras that don't teach, including background stories and decorative sound.

If you use text-to-speech, choose a modern neural voice and listen to a full lesson before you publish. In Kveeky you can choose from 700+ voices in 40+ languages and adjust speed, volume and emotion. See our AI voiceover for e-learning page for course narration samples.

For the practical side, see when to use more than one narrator in a course. Also read how narration quality affects course drop-off, and how to make course audio and video accessible.

Frequently asked questions

Does listening improve learning more than reading?

It depends on what else is on screen. With an animation or diagram, narration beat on-screen text in Mayer and Moreno (1998). With no other visual, text plus narration helped in Moreno and Mayer (2002).

Should on-screen text match the narration word for word?

Not when a diagram or animation is also on screen. Mayer, Heiser and Lonn (2001) found duplicated on-screen text lowered retention and transfer. Use short labels instead.

Are AI voices as effective as human voices for learning?

Modern ones were close in recent studies. Craig and Schroeder (2019) found a modern text-to-speech engine was not statistically different from a human voice in most respects. Older synthetic voices did worse in a 2003 study.

What is the voice principle in multimedia learning?

It's the finding that people learn better from a human-sounding voice than from a machine-like one. It comes from studies such as Mayer, Sobko and Mautone (2003), though newer work suggests modern text-to-speech has narrowed the gap.

Does a conversational tone improve retention?

In Mayer and colleagues (2004), conversational wording improved transfer, meaning applying the ideas, but not retention. It's still worth using, because applying knowledge is usually the goal of training.

How we checked this guide

This guide is written by Ankit Agarwal for the Kveeky team. Disclosure: Kveeky makes an AI voice generator. We rewrote an older version of this post and kept only claims we could trace to a published study.

Studies cited (abstracts read October 6, 2026):

• Mousavi, Low and Sweller (1995), Journal of Educational Psychology 87(2): doi.org/10.1037/0022-0663.87.2.319
• Mayer and Moreno (1998), Journal of Educational Psychology 90(2): doi.org/10.1037/0022-0663.90.2.312
• Mayer, Heiser and Lonn (2001), Journal of Educational Psychology 93(1): doi.org/10.1037/0022-0663.93.1.187
• Mayer and Chandler (2001), Journal of Educational Psychology 93(2): doi.org/10.1037/0022-0663.93.2.390
• Moreno and Mayer (2002), Journal of Educational Psychology 94(1): doi.org/10.1037/0022-0663.94.1.156
• Mayer, Sobko and Mautone (2003), Journal of Educational Psychology 95(2): doi.org/10.1037/0022-0663.95.2.419
• Mayer, Fennell, Farmer and Campbell (2004), Journal of Educational Psychology 96(2): doi.org/10.1037/0022-0663.96.2.389
• Craig and Schroeder (2019), Journal of Educational Computing Research 57(6): doi.org/10.1177/0735633118802877
• Chiou, Schroeder and Craig (2020), Computers & Education 146: doi.org/10.1016/j.compedu.2019.103756

No Kveeky usage data is used in this guide. Kveeky facts come from kveeky.com/pricing.

Want to test this on your own course? Rewrite one formal lesson in a conversational style, generate it with a modern voice, and compare quiz results. Our hub on how businesses use text-to-speech tools covers other training and marketing uses.