AI Voices for Audiobooks and E-learning in Malay and English: A Practical Guide

Short answer: Long-form audio is a different job from a 30-second advert. Write the script for the ear, split it into short sections, pick one voice by listening to your real text, fix the settings, and then check every join. Publish a transcript with the audio, confirm you hold the rights to the text and the voice, and expect each regenerated section to sound similar, not identical.

This guide covers audiobooks and e-learning modules in Malay and English. The examples come from voiceover.my demos, where every clip shows its exact transcript, voice and request.

What changes when people listen for longer

An advert has to work for half a minute. A book chapter or a training module has to work for ten, twenty or sixty minutes, often while the listener is doing something else. Small problems that pass unnoticed in a short clip become tiring over a long run.

Qualitative guidance by voiceover.my, 3 November 2026, based on common production practice and our own tests. It is not a published standard.

Three of these deserve attention before you write a word. Pace: listeners cannot skim audio, so long-form narration needs a steadier pace and short gaps to take in a new idea, a scene change or a slide. Consistency: a book is one reader and a course is usually one presenter; if chapter four sounds brighter or faster than chapter three, listeners notice. That is mostly a process problem, solved by fixed settings, short sections and careful listening at every join. Names and terms: a misread name repeats on every page. Malay courses often carry English loanwords and acronyms, and English content for Malaysian learners often carries Malay names and places. Test them all first.

How pace looks in our demo clips

The voiceover.my gallery has 8 audiobook clips and 8 e-learning clips, each set split evenly between Malay and English, as of 3 November 2026. Four of the 16 are from ElevenLabs and the rest from Google Gemini. Every audiobook clip reads the same chapter opening in its language, and every e-learning clip reads the same short budgeting lesson.

Calculated by us from the clip durations (rounded to whole seconds) and transcript word counts shown on each demo page, 3 November 2026. The Google clips used Gemini 3.8 Flash-Lite TTS with the style lines warm, unhurried storytelling (audiobook) and clear, encouraging, steady pace (e-learning). The ElevenLabs clips (Cikgu Aisyah and Matilda) used eleven_v4 with no style sent. Four clips per row is a small sample, so treat these as a planning reference, not a rule.

Word counts mislead across languages. The Malay and English audiobook scripts are close in length when you count characters (about 380 and 395), but the Malay one has 59 words against 71. In these scripts the Malay words are longer on average, so a words-per-minute target from English content will not transfer. Plan Malay timing by characters, or simply time a test section.

The audiobook clips ran slower than the e-learning clips in both languages. The scripts differ, so this is not a controlled comparison, but it fits what the style lines asked for. If narration feels rushed, test a calmer style line against a no-style control.

You can hear them side by side on the audiobook demos and e-learning demos pages. Good starting points:

• Book chapter opening, Malay, Despina (Gemini)
• Book chapter opening, Malay, Cikgu Aisyah (ElevenLabs)
• Chapter one, English, Charon (Gemini)
• Monthly budget lesson, Malay, Kore (Gemini)
• Monthly budget basics, English, Schedar (Gemini)
• Monthly budget basics, English, Matilda (ElevenLabs)

On each page, As requested shows what was asked for or reported by the provider (model, voice, style). Human-reviewed shows what a person heard when listening, such as “Female voice” or “Fitting for this scene”. They are not the same thing, and our reviews come from a single listener, not a panel.

How to write a script for TTS

A text written for the page is not automatically a good script. These habits help in both Malay and English.

• Write the words exactly as they should be spoken. TTS reads what you give it. Google describes Gemini TTS as tailored for exact text recitation, with audiobooks as one example. That is what you want for a book, but it means the script must already be right.
• Keep sentences short to medium. Long sentences with several clauses are hard to follow by ear, whoever reads them.
• Spell out what should be said. Decide how numbers, dates, currencies and acronyms are spoken (“RM” as “ringgit”? “KPI” as letters?) and write them that way.
• Use signposts in e-learning. Our e-learning demo script uses “Mula-mula”, “Kemudian” and “Akhir sekali” in Malay, and “First”, “Next” and “Finally” in English. Signposts help learners keep their place.
• Mark chapter and section titles clearly. Our audiobook script opens with “Bab satu.” and “Chapter one.” as separate short sentences.
• Plan pauses. Google’s documentation describes inline tags such as <short pause> placed directly in the transcript. In our own tests, <short pause> is one of the tags we have checked by listening. Paragraph breaks and full stops also help. We have not tested inline tags with ElevenLabs.

A pronunciation list

Before you generate anything long, build a short list of every name, term and acronym that repeats. For each one:

1. Put it in a test sentence that looks like the real text.
2. Generate it with your chosen voice and listen.
3. If it is wrong, try a respelling in the script (for example, splitting syllables or spelling it as it sounds) and test again.
4. Record the spelling that worked and use it everywhere.

Keep the correct spelling too: the transcript needs it, not the respelling.

Chunking long text

Do not try to generate a whole book or course in one go. Both providers set per-request limits:

• Google lists an input limit of 8,192 tokens and an output limit of 16,384 tokens for both Gemini 3.8 TTS models. Google’s pricing page says audio tokens correspond to 25 tokens per second of audio, which by our arithmetic puts the output cap at roughly 11 minutes per request.
• ElevenLabs lists a 10,000 character limit for eleven_v4, approximately 10 minutes of audio, and suggests splitting longer content into multiple requests.

Staying well under those limits is a good idea anyway. Short sections are easier to check, cheaper to regenerate and easier to replace when the text changes.

Working guidance by voiceover.my, 3 November 2026. Adjust section size to your editing tool and review process.

Split at paragraph ends, scene breaks or slide changes, never mid-sentence. Name files so they sort in order (for example ch03-s04.wav) and log the text, voice, model and style behind each one.

Google’s documentation notes that single requests return WAV with a header by default, and that you can request raw PCM if you plan to join pieces directly. Leave a short, consistent gap between sections.

Choosing a voice by listening

Labels such as “warm” or “storyteller” are a starting point, not a decision. Google’s documentation shows that its voice list can be filtered by usage context, including “Audiobook”, which is a useful way to build a shortlist. After that, only listening counts.

1. Pick a test passage of 2 to 3 minutes from the real text. Include dialogue, a hard name and a number.
2. Generate it with three or four voices, same model and settings, plus a version with no style instruction as a control.
3. Listen to the whole passage, not the first ten seconds. Stamina matters more in long-form audio than in any other use.
4. Listen on the devices your audience uses. Earbuds on a commute, a laptop in a training room, a phone speaker.
5. Ask people from your audience to listen blind. For Malay content, Malaysian listeners can tell you what one reviewer cannot. A provider supporting “Standard Malay” or “Malay” does not prove the result sounds local.
6. Fix the choice. Once chosen, do not change voice, model or style mid-project.

For the scorecard and the one-variable-at-a-time method we use, see how to choose an AI voice.

Audience and voice age

We only use adult voices, and we do not publish clips that sound like a child. For children’s books and courses, a cheerful, friendly adult voice is our approach.

Consistency and regeneration: what to expect

AI speech is not a recording you can replay. On our demo pages we state it plainly: re-creating a request gives a similar clip, not an identical one, and output varies between runs. That has practical effects for long-form work.

• Lock every setting you can. Same voice, same model, same style line, same output format, for the whole project. Record them in your log.
• Regenerate by section, then check the joins. A replaced paragraph may differ slightly in pace, energy or loudness from its neighbours. Listen from the end of the previous section into the start of the next.
• Keep the takes you approve. Store the approved audio file, not just the recipe. You cannot get the exact same file back later.
• Watch voice availability. ElevenLabs Voice Library voices are shared by their creators and, as our demo pages note, can be renamed or withdrawn. Keep your approved files and check the voice before starting a new volume.

Google positions Gemini 3.8 Flash TTS for long-form narration that needs stable voice and room tone, and Flash-Lite as the faster, cost-efficient option. ElevenLabs describes eleven_v4 as suited to audiobook production. Our own demos used Flash-Lite and eleven_v4 and are short, so they say nothing about stability over a whole book. Test on your own text.

Google Gemini or ElevenLabs?

Last verified 3 November 2026 on the official pages listed under Sources. The last row describes our gallery, not each provider’s full catalogue. We do not rank the providers: our demos are short and were judged by one listener.

On price, Google’s pricing page lists standard paid audio output at $9.00 per million tokens for Gemini 3.8 Flash TTS and $6.00 for Flash-Lite through 31 December 2026, doubling to $18.00 and $12.00 from 1 January 2027 (last verified 3 November 2026). For ElevenLabs, check the official pricing page for current plans and credits. For a long book, budget for retakes as well as the final length. Our pricing reference shows the rates we use for planning.

Accessibility basics

Audio-only learning shuts out people who cannot hear it, and some who process text better than speech. The W3C’s Web Accessibility Initiative (WAI) is clear on the basics:

• Transcripts. WAI describes a basic transcript as a text version of the speech and non-speech audio information needed to understand the content. It lists people who are Deaf, hard of hearing or have difficulty processing auditory information among those who use them.
• Audio-only content. For prerecorded audio-only media such as a podcast, WAI says transcripts are required at WCAG Level A.
• Video with audio. For prerecorded video with audio, which covers most e-learning videos, WAI says captions are required at WCAG Level A and transcripts at Level AAA.

A TTS project has an advantage here: you already have the script. Use it as the base for the transcript or caption file, check it against the final audio, add any important non-speech information, and use correct spellings rather than respellings.

Rights and licensing: questions to ask first

This section is general caution, not legal advice. For a commercial book or course, get advice from someone qualified in your jurisdiction.

• Do you hold the audio rights to the text? Owning a printed edition or an e-book licence is not the same as the right to publish an audio version. Check with the author or publisher.
• What does the provider’s licence allow? Google’s Gemini API terms say Google won’t claim ownership of generated content, that it may generate the same or similar content for others, and that you are responsible for your use of it. ElevenLabs’ pricing page lists a commercial licence from its Starter plan upward. Read the current terms for the plan you actually use.
• Who owns the voice? A library voice shared by another creator is used under that provider’s terms and may be withdrawn. Do not clone or imitate a real person’s voice without their clear consent.
• What do your distribution platforms allow? Audiobook stores and learning platforms set their own rules on AI narration and disclosure, and those rules change. Read them before you produce the full work.

The long-form checklist

• I hold the audio rights to the text.
• The script is written for the ear, with numbers, dates and acronyms spelled out.
• I have a pronunciation list and tested every entry.
• The text is split into short sections at natural boundaries.
• I chose the voice from a 2 to 3 minute test passage, not a preview line.
• People from my audience heard the shortlist blind.
• Voice, model, style and format are fixed and logged for the whole project.
• I listened to every join between sections.
• I kept the approved audio files, not just the settings.
• A checked transcript or captions go out with the audio.
• I read the provider’s and platform’s current terms.

Ready to compare voices? Start with the full gallery and filter by situation and language.

For the background on Malay voices, accent and cost, read our complete guide to Malay text-to-speech. If you are deciding between AI narration and a person, see AI voice-over versus a human voice actor in Malaysia.

Sources

All sources below were accessed on 3 November 2026.

• Speech generation (TTS), Gemini API documentation, for exact text recitation, inline tags, output formats, voice filtering by context, model positioning and Standard Malay support.
• Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, for input and output token limits.
• Gemini Developer API pricing, for audio output rates, the 1 January 2027 change and 25 audio tokens per second.
• Gemini API Additional Terms of Service, for the use of generated content.
• ElevenLabs models overview, for eleven_v4 positioning, Malay support and the character limit.
• ElevenLabs pricing, for the plan structure and commercial licence.
• Transcripts, Making Audio and Video Media Accessible (W3C WAI), for transcript and caption requirements.