How to Turn Text into Audio (Free, Audiobook, iPhone)

The short answer

If you just need text spoken aloud right now — to proofread your own writing, to follow a recipe, to listen to an article while commuting — your operating system already ships a speech engine and nothing you install will beat its convenience. On every major platform the read-aloud feature is one gesture away and costs nothing.

If you need an audio file you can keep, embed, or distribute, you need a synthesiser that renders to a file rather than a voice that only plays through the system speaker. A neural text-to-speech model running in your browser does exactly that: it converts your text into a downloadable WAV. No account, no upload, no API key.

If you need narration that passes for a human voice — a product demo, a kiosk prompt, a chapter of a book you are distributing — you need that same neural model, but you also need to write the source text for it and to check the output before publishing. Machine voices get close enough that the seams are specific and fixable: wrong pauses, mispronounced names, and flat intonation at sentence endings. The craft sections further down address each one.

• Listening to text: your OS already does it. Free, instant, nothing to install.
• A reusable audio file: a neural model that renders to WAV in your browser.
• Narration that sounds human: the same model plus writing and checking habits.

Three things you might actually want

Most people who search “turn text into audio” have not said out loud which of three outcomes they need. The engine choice, the export format, and how much effort to invest in the writing all change depending on the answer, so sorting this out first saves time.

Listening to text

This is the accessibility and convenience case. You want words spoken so you can verify a draft by ear, follow a recipe with flour on your hands, or listen to a long article while walking. You do not need a file, you do not need a specific voice, and you do not need to install anything.

Every major operating system ships a speech synthesiser and a read-aloud shortcut. macOS has Spoken Content under Accessibility. Windows has Narrator and a “Read Aloud” button in Edge. iOS and Android both offer system-wide selection-speak. These use the operating system's built-in voice, they are genuinely free, and for listening purposes they are as good as anything you will set up in the next five minutes.

Nothing you install beats their convenience here. There is no reason to download a model, open a browser tab, or paste text into a form when the OS can read the paragraph you just highlighted.

A reusable audio file

This is the case where the built-in voice is not enough. You need something you can hand to a video editor, upload to a phone tree, embed in a learning module, or distribute as part of an audiobook. The key property is “renders to a file you own,” and here the difference between engine types matters.

The operating system's speech engine is designed to play through the system audio path. Some browsers expose a way to capture that output, but it is inconsistent, and the system voice itself is often not licensed for redistribution as a file. Even if you manage to record it, you are working against the OS's assumptions rather than with them.

A neural text-to-speech model that renders to a file gives you something you can keep, move, and ship. The browser tool linked at the bottom of this page does exactly that: you paste or type text, the model synthesises it on your device, and you download a WAV file. Batch mode lets you generate several passages in one session rather than one at a time.

Something that sounds like a person

This is the hardest goal, and it is where the writing craft and the pre-publish checks in the sections below actually matter. Modern neural TTS is far closer to natural speech than the old robotic formant synthesizers, and for short passages, a clean WAV from a model like Kokoro can pass as recorded narration in informal contexts. The seams show up in longer material and in specific, predictable places — mispronounced names, flat questions, a drift in pacing over many paragraphs — and each of those has a countermeasure.

One honest boundary: the English page's model (Kokoro v1.0) offers four voices — two American (male and female) and two British (male and female). It is an English-only model. If you need Spanish, Portuguese, Indonesian, or any other language, the neural model here will not help, and you should look at OS voices for that language or a service that specifically supports it.

What actually makes a voice sound natural

People say “the voice sounds robotic” and assume the problem is timbre — the texture of the sound itself. Modern neural models have largely solved timbre. What still makes a synthesis sound mechanical is three prosodic features that are harder to model.

Pitch contour is the shape of the melody line across a sentence. Natural speech rises at the start of a clause, peaks at a stressed syllable, and falls at the end of a declarative sentence. A question rises sharply before the final word. A sentence that should be read with surprise or doubt bends differently again. When a model flattens this contour — when every sentence ends on the same falling note regardless of punctuation — the result sounds generic, even if the individual phonemes are correct.

Phrasing groups words into breath units the way a speaker would. “The cat sat on / the mat” is one unit. A poorly phrased synthesis groups “The cat / sat on the / mat,” splitting at places no human speaker would. The listener registers that as “reading aloud” rather than “telling me something.” Models trained on large speech corpora learn common phrasing boundaries, but they struggle when a sentence has unusual structure, long subordinate clauses, or punctuation the model saw infrequently in training.

Pause placement is the third axis. Comma = short pause. Period = medium pause. Paragraph break = long pause. A good model respects all three, but in practice many engines treat a paragraph break as “the same silence as a period” unless explicitly configured, which makes narration feel breathless — like a reader who cannot stop for air.

Long-form narration is where these three problems compound. Over a single sentence, even a mediocre model sounds acceptable. Over 200 sentences, three specific things break the illusion:

• Drift: the tempo or cadence subtly changes across a long passage — the first paragraphs sound measured, the last paragraphs sound rushed or unnaturally slowed.
• Mispronunciation: proper nouns, abbreviations, and unusual vocabulary are handled by the model's learned phoneme mapping; when a word is rare in training data the synthesis guesses wrong, and it guesses wrong consistently so every re-generation makes the same mistake.
• Flat question intonation: a question mark at the end of a paragraph or a section boundary often loses its rising contour because the model is also trying to signal the structural close.

How to write for a voice

The most common mistake people make with TTS is feeding it text written for the eye. Web copy, academic writing, and even email prose are structured for scanning: long subordinate chains, embedded parentheses, abbreviations, dense numbers. None of those translate cleanly into audio, and the model has no way to know you meant them visually.

The fix is to write or rewrite specifically for the ear. It is not hard, and once you build the habits the first synthesis sounds dramatically better without any post-production.

One thought per sentence

Keep sentences short. Aim for 15 to 20 words. A model that is reading a 45-word sentence with two subordinate clauses has to guess where the main verb lands and will often flatten the pitch arc on the subordinate part, making the whole thing sound uncertain. Short declarative sentences give the model a clean contour to follow.

If you have long sentences in your source text, break them into two or three shorter ones before feeding the text in. You lose nothing in audio; you gain clarity.

Spell out abbreviations

Write “Doctor” instead of “Dr.” if you cannot verify the model expands it correctly. Write “United States” instead of “U.S.” if the model might say “ooh-ess” or “yoo-ess.” Write “N A S A” if you want letter-by-letter pronunciation, or “the space agency” if you want it read as a word. The model does not know which convention you intend; you do.

For abbreviations that are commonly read as words (“NASA,” “FBI,” “UN”), test them once. If the model gets it right every time, keep the abbreviation. If it does not, spell out or respell.

Write numbers the way they should be said

The number “2026” is three different readings: “two thousand twenty-six,” “twenty twenty-six,” and “two-oh-two-six.” The model picks one based on its training distribution and the surrounding context. If you want a specific pronunciation, write it that way: “twenty twenty-six.” This is especially important for phone numbers, addresses, years in historical narration, and quantities where “a thousand and two” versus “one thousand two” changes meaning.

Decimals: “3.14” might be “three point one four” or “three fourteen.” Write “three point one four” if the context does not disambiguate.

Use punctuation as a pause instruction

A period tells the model “this thought is over; take a breath.” A comma says “pause briefly.” A paragraph break says “this section is finished; take a longer breath.” Some engines treat all three the same, but neural models trained on natural speech data distinguish them.

If the model rushes through a transition where you want a pause, try an em-dash or an ellipsis. “The answer was clear — it did not work” gives the model a beat before the reveal. “The answer was clear... it did not work” might give an even longer one, depending on the engine. Test both.

If you want a pause where there is no punctuation in the original, add one. Write “And then, after a long silence, she spoke” rather than “And then after a long silence she spoke.” The extra commas give the model permission to slow down.

Respell names phonetically

If the model reads “Siobhan” as “See-oh-ban” and you know it will not learn the correct pronunciation from context, respell it in the input: “Shih-VAWN.” This looks wrong on the page but sounds right in the audio. You can also use the spelling a less-trained reader would guess correctly: “Shivawn,” “Shivaun.”

Proper nouns from fiction are where this hurts most — invented names from fantasy novels, sci-fi terms, character names from non-English cultures. The model has never seen them; it falls back to grapheme-to-phoneme rules that may produce something close but wrong. Spell them the way you want them said.

Generate in sections, then stitch

The instinct when you have a long document is to paste the whole thing into the text box and hit Generate. This is almost always the wrong approach, and the reasons are practical rather than technical.

When something goes wrong in a 3,000-word generation, the only fix is to re-generate the entire passage. If the model mispronounces a character's name in paragraph four, you either accept it or re-render everything. If the pacing drifts by paragraph twenty, you have no way to fix just that section.

If you generate in sections, you can isolate problems. One bad section means one re-generation. A name that keeps coming out wrong means respelling it in one or two section inputs rather than throughout the whole file. Drift on a long passage means regenerating the last section with the corrected name already in mind.

Break at natural boundaries: chapter endings, section headings, a blank line between topics. Do not break mid-sentence unless you have no choice — the model needs to see the sentence arc to get the intonation right, and if you cut a sentence in half the second clip will start with an unnatural pitch.

Keep the model, voice, and any speed setting the same across every section. If you change the voice parameter between parts, the stitched file will have an audible timbre shift at every boundary.

To join the pieces, use any audio editor — even a free one — and concatenate the WAV segments end to end. If the transitions sound abrupt, crossfade a few frames. If the model already inserted end-of-sentence silence, the join will be clean enough that you do not need to do anything. The browser tool's batch mode helps here: you can feed multiple passages and generate them all in one session, which keeps the settings consistent.

Before you publish: three things to check

You do not need to listen to every word of a generated file before shipping it, but you do need to check three specific things. They are the mistakes that a machine voice cannot tell it is making, and a listener catches instantly.

The first and last second. Listen to the very beginning and the very end of every clip. Two common failure modes: the model clips the first phoneme (so “The” becomes “he”), and the tail gets truncated or has dead silence. Both are obvious once you know to check. If the start is clipped, regenerate with a brief pause or silence prepended in the text. If the end is cut, make sure the input ends with a period or sentence-final punctuation so the model knows the utterance is complete.

Every number. For each digit sequence in your input, listen to how the model rendered it. “2026” might be “two thousand twenty-six” when you needed “twenty twenty-six.” A phone number might be grouped wrong — “five five five, one two three four” versus “five fifty-five, one twenty-three four.” If you have multiple numbers, spot-check each one because the model's behaviour is context-dependent, not uniform.

Every proper noun. Character names, city names, brand names, scientific terms, any word that appears only in your text and nowhere in common training data. Listen for whether the model guessed right. If it consistently gets “Nwoye” or “Gilead” wrong in a novel excerpt, respell it in the input text and re-generate that section.

These three checks catch almost every publication-embarrassing moment. The rest — general pacing, timbre consistency, whether the reading sounds like it understands the material — comes from how you wrote the input text, not from post-processing the audio.

Audiobook narration: what to know before you record

If you are searching “turn text into audio for audiobook,” the honest answer is: it depends on where you want to distribute it, and that answer changes.

Narrated books sold through major retail channels have their own quality standards and approval processes. The bar is not “does the voice sound reasonably human?” — it is a longer checklist about consistency across the whole book, room tone, noise floor, editing standards, and listener expectations shaped by human narrators who can act. These channels do not currently accept every machine-narrated title, and their policies on synthetic voice have evolved.

What you need to do before you spend hours generating a full book: read the distributor's current policy on machine or synthetic narration. Not last year's blog post — the policy that is live when you are ready to submit. That document will tell you whether they accept TTS audiobooks at all, whether they require specific audio technical standards, whether there is a quality review that may reject a narration that meets the specs but sounds unnatural, and whether the content category matters (fiction vs. non-fiction, public domain vs. newly published).

Do not assume any channel accepts machine narration just because the technology has gotten better. And do not assume it rejects it either — some categories and some distributors do accept it under specific conditions. The only way to know is to read the policy before you record.

If the distributor's policy permits TTS, and you clear the quality bar, the same craft sections above apply: write for the voice, generate in sections, check every name and number. Long-form fiction is the hardest case because drift compounds over hours and because listeners have strong expectations about character voice consistency across an entire book.

On iPhone and Android

People search “turn text into audio on iPhone” and the answer is general knowledge, not this browser tool. Mobile operating systems ship a speech synthesiser at the system level, and they ship accessibility features that read selected text aloud.

On iOS, the system includes a speech engine accessible through the Accessibility settings; selected text can be read aloud in apps that support it, and the Shortcuts app can produce audio from text as an automation step. On Android, the system TTS engine is accessible from any app that requests it, and the built-in Google Text-to-Speech component can speak selected text if you enable that in Accessibility settings.

What you can export from these mobile routes varies. The accessibility read-aloud features are designed to play through the speaker, not to produce a file. If you need a saved audio file, you either need an app that specifically exports (and its capabilities depend on what that app does), or you can use a browser-based tool like the one linked at the bottom of this page, which works on mobile Safari and Chrome just as it does on a desktop — you get the same WAV download.

Do not trust a guide that claims a specific iOS version's menu path or a specific export format as though that interface will never change. Apple and Google both reorganise their settings between releases. What is stable is the principle: both platforms can speak text, neither was designed to be a TTS-file-production tool, and if you need a file the browser route gives you one.

The browser tool: what it does and what it gives you

The ToolHubNest text-to-speech page runs a neural model called Kokoro v1.0 entirely inside your browser tab. Here is what that means concretely, and what it means for the file you get out.

The model produces speech by converting your text into a waveform on your own processor. Your text does not go anywhere — there is no server receiving it, no queue, no API call. The page generates a WAV file and hands it to your browser as a download.

The model's voices are English. Four options: US English female, US English male, UK English female, UK English male. That is the boundary; the model is English-only and it is not going to speak Spanish or any other language. If you need another language, use your operating system's voices or find a tool that explicitly supports the one you need.

Output format: WAV. This is uncompressed, which means the file is larger than an MP3 of the same audio, but it is also the highest fidelity you can get from the model — no lossy encoding artifacts baked in before you start editing. The page does not offer MP3 export. If you need MP3 for a specific player or platform, convert the WAV in any audio editor after download.

Batch mode lets you generate more than one passage in a session. Instead of pasting text, hitting Generate, downloading, repeating — you can queue multiple passages and get a set of WAVs from one run. This is useful when you are narrating chapters, building a set of prompts for an IVR tree, or creating audio for multiple slides in a presentation.

The model weights download on first use. Your browser reuses the cached copy on later visits, so the first run in a new browser pulls a larger payload and subsequent runs go straight to synthesis. If the model cannot load (network issue, unsupported browser), the page falls back to the browser's own speechSynthesis engine — which requires no download but does not produce a WAV file and is limited to the voices your browser has available.

What is not claimed: any accuracy score, any latency figure, any model-size number, any guarantee that the page functions with no network at all after the first load (the page loads fonts and runtime scripts from CDNs), or any indication that your results are stored. This is a one-shot tool: generate, download, move on.

• Model: Kokoro v1.0, neural TTS, English-only, runs in-browser on your CPU.
• Voices: US English (male/female) and UK English (male/female).
• Output: WAV download. No MP3 export. No upload.
• Batch mode: generate multiple passages in one session.
• First use downloads model weights; subsequent uses rely on browser cache.
• Fallback: browser speechSynthesis if the model cannot load.

Frequently asked questions