How to convert text to speech with AI, step by step

You already use it when your maps app tells you to take the next exit. Text to speech (TTS) is any system that reads written text aloud and hands you the sound.

Older voices were built by stitching together snippets of recordings or computing the sound from fixed rules. Hence the flat tone and choppy syllables: the system didn't understand the sentence, it just sounded it out.

Neural voices learn to speak from thousands of hours of real people and generate the waveform from scratch. The jump was measured with WaveNet. In Google's tests, listeners rated its US English voices 4.1 out of 5, more than 20% better than standard voices, closing the gap with human speech by over 70%. Since then, the model decides where to breathe and which word to stress.

So whether a tool sounds human is no longer the question. The real differences are more mundane: how much text it takes, whether you can download the MP3, whether you can use it commercially and what each minute costs.

Before you paste your script into the first free site that shows up, check these six things. You find out about the last four too late: the video's already edited, and it turns out you can't download the audio or can't monetize it.

We compare per minute of audio, using the cheapest plan you can actually buy from each service. Where a price is published in characters, we convert it at our measured average of 1,000 characters per minute.

Watch the last column. On a subscription, what you don't use in a month is usually gone. With a one-off purchase, it stays.

Here's how to turn text into a downloadable MP3 with nothing to install:

1. 1Open the generator. Paste your text or upload a document (.txt, .md, .srt, .pdf, .docx). A single file can hold up to 600,000 characters, roughly 10 hours of speech.
2. 2Pick one of 300+ voices, filtering by language, accent, gender or style. Play the sample first, and if you're torn between two, generate one paragraph with each.
3. 3The language is detected from your text, in any of the 7 supported: English, Spanish, German, French, Portuguese, Italian and Russian.
4. 4Adjust the speed (0.5× to 2×), the pause between sentences (up to 800 ms) and the volume. You can preview the settings before generating.
5. 5Click Generate. You see the cost up front: 60 credits per 1,000 characters (about 60 per minute of audio), with a minimum of 100 per file.
6. 6Listen back and download the MP3. Save it to your device: files stay in your history for 4 days.

The voice reads exactly what's on the page, punctuation included. Punctuation is how you direct it, and anything that's ambiguous on paper, like a date or an abbreviation, can come out in a way you didn't expect.

The voice shapes the result more than any setting. Think about who's listening and for how long: a high-energy voice sells a 30-second ad and wears people out over a six-hour audiobook.

On average, one minute of audio is about 1,000 characters including spaces. But the voice matters more than the text: the same script runs almost 50% longer with the slowest voice we've measured than with the fastest.

Voizum delivers mono MP3 at 96 kbps, which opens in any player or video editor. A voice only needs one channel, and at that bitrate there's no audible loss in speech. Each minute is about 0.72 MB; an hour, roughly 43 MB.

If a platform asks for more, exporting at 192 kbps meets the spec, but it still sounds like the MP3 you started with.

On Voizum, the audio you generate is yours to use commercially: monetized videos, ads, courses or podcasts, with no credit to the tool required. Two limits apply: you can't use it to impersonate someone or for illegal content, and you can't clone someone else's voice without their permission.

And here's what the platforms ask:

• YouTube doesn't require you to disclose an AI voiceover or dub. It does require labeling realistic content that makes a real person appear to say something they didn't.
• To earn from ads, YouTube excludes "inauthentic" content: mass-produced, repetitive videos with no original input. The AI voice is fine; what gets a channel turned down is a recycled script or a template repeated a thousand times.
• Audible, through ACX, requires audiobooks to be narrated by a human unless they're part of its own authorized synthetic voice programs. If Audible is your target, confirm this before producing the whole book.

To hear a text once, you don't need a website at all. Microsoft Edge has a free Read Aloud with natural voices, and your operating system has a built-in reader. The catch: they read live and don't give you a file.

If you need the MP3, sign in with Google and you get 600 free credits every month, about 10 minutes of audio, to use with any voice in the library. They renew automatically, don't roll over and need no card.

If you need more, credits are bought with no subscription and never expire. The free monthly credits are for generating audio on the website; voice cloning, the API and batch generation use purchased credits.

The ones you hear first:

• Pasting text without cleaning it up: stray line breaks, page numbers or leftover formatting create odd pauses or get read out.
• Picking a voice from its 10-second demo without testing a real paragraph of your script.
• Leaving acronyms, abbreviations and ambiguous dates (10/2/2026) the way a document shows them instead of the way people say them.
• Four-line sentences: fine on paper, but in audio the listener loses the subject.
• Generating a long script in one go without listening to a minute first: if the voice or the speed is off, you pay for all of it.
• Not downloading the MP3: files are removed from your history after 4 days.