AI Text to Speech Voiceover - MP3, WAV, Opus and AAC

Turn a script, or a whole batch of them, into a text to speech voiceover you can download: one MP3, WAV, Opus or AAC file per script, in one of six voices, at the speed you set. Each file lands in the run's key-value store with a direct link, next to a row giving the voice, the length and how many characters were read.

The part to know before you start: this runs on your own OpenAI key. You paste the key in, OpenAI bills you for the speech itself, and without a key the run hands back a free demo record instead of audio.

Input	A script in text, or a list of them in texts
Output	One audio file per script, and a row with its link, length, voice and character count
Ceiling	No item cap. Long scripts are split and joined back, so the run's timeout is the real limit
Account needed	Your own OpenAI API key. Without one you get a free demo record
Price	$40.00 per 1,000 voiceovers, so 4 cents each, flat on every plan. The free plan's $5 a month covers about 125. OpenAI bills your key for the speech separately

🔍 What AI Text to Speech Voiceover does

It reads your text out loud and hands you the file. Six voices, four formats, and a speed dial from 0.25 to 4.0.

A long script does not need splitting by hand. Anything over about 3,500 characters is cut at sentence boundaries, each piece is voiced on its own, and the pieces are joined back into a single file without re-encoding. The row tells you how many pieces went into it, in chunks.

texts is the batch door. Give it an array of strings and you get one file per string. Give it an array of objects and it reads whichever of script, scriptText, text or narration is present. An object with none of those four keys is dropped quietly, which is worth knowing if a batch comes back shorter than you sent.

Items are done one after another rather than all at once, so a long list takes a while.

📋 What data you get from each voiceover

What you get	Field
The audio file, saved in the run's key-value store	audioKey
A direct link to that file	audioUrl
Its length in seconds, measured off the finished file	durationSeconds
The voice, model and format actually used, after defaults	voice, model, format
How long the text was, and how many pieces it was voiced in	characters, chunks
The first 160 characters of what was read, so rows are easy to tell apart	textPreview
Which item of your batch it was, counting from 1	index

▶️ How to turn text into speech

1. Open AI Text to Speech Voiceover and click Try for free.
2. Paste your OpenAI key into OpenAI API key (BYO).
3. Put your script in Text / script, or a list of scripts in Texts (batch).
4. Pick the Voice and the Audio format, then click Start.
5. Take the files from the run's Storage tab, or follow audioUrl from each dataset row.

Start it once with no key and you get the free demo record, so you can see the shape first.

💰 How much does it cost to turn text into speech?

$0.04 per voiceover, which is $40.00 per 1,000. Flat on every Apify plan, no volume tiers. Fifty narration blocks in one batch run come to $2.00, and on the free plan the $5 Apify gives you each month covers about 125 voiceovers.

You pay per delivered file. The keyless demo record and any item that failed are not charged.

OpenAI bills your own key separately for the speech itself, and tts-1-hd costs you more there than tts-1.

📥 What you give it

{
  "text": "Welcome back to the channel. Today we are looking at one of the strangest mysteries of the deep ocean.",
  "voice": "onyx",
  "format": "mp3",
  "speed": "1.0"
}

Field	Default	What it is
text	none	One script, as a single string. The form opens with an example sentence filled in, which is a starting point rather than a default.
texts	none	Batch mode. An array of strings, or of objects keyed script, scriptText, text or narration. One audio file per item.
voice	onyx	One of alloy, echo, fable, onyx, nova, shimmer. onyx is the deep male one, nova and shimmer are female.
model	tts-1	tts-1 is quick, tts-1-hd sounds better and costs you more on your own OpenAI bill.
format	mp3	mp3, wav, opus or aac.
speed	1.0	Playback speed between 0.25 and 4.0. It is a text field, so type a number: anything that is not one quietly becomes 1.0.
openaiApiKey	none	Your own OpenAI key. Marked secret, so it is not written into the run's visible input.
baseUrl	none	Advanced. Any OpenAI-compatible /audio/speech host. Left empty it uses https://api.openai.com/v1.

Give it a key but no text and no texts and the run fails, because there is nothing to read.

📤 What you get back

One dataset row per finished voiceover, and the audio itself in the run's key-value store under a key like voiceover-1-1757000000000.mp3.

No run on record has delivered a voiceover with a customer's key, so rather than type out a plausible row, here is the real demo record a keyless run wrote on 2 October 2026, notice cut short. A delivered row has the same fields, with audioKey, audioUrl and a real durationSeconds filled in, and no _demo or _notice.

{
  "_demo": true,
  "_notice": "SAMPLE OUTPUT ... Supply \"openaiApiKey\" to generate real voiceover audio.",
  "ok": true,
  "index": 1,
  "voice": "onyx",
  "model": "tts-1",
  "format": "mp3",
  "characters": 103,
  "chunks": 1,
  "durationSeconds": 0,
  "audioKey": null,
  "audioUrl": null,
  "textPreview": "Welcome back to the channel. Today, we're diving into one of the strangest mysteries of the deep ocean."
}

Field	How to read it
ok	true on a delivered voiceover.
chunks	How many separate pieces were voiced and joined. 1 means the script fitted in one.
durationSeconds	Measured off the finished file. It reads 0 when that measurement failed, not when the audio is empty.
audioUrl	A direct link to the file. audioKey is its name in the key-value store.
characters	The length of the text you sent for that item.

🧾 Reading the output

Three kinds of record come out of a run, and only one of them is in the dataset.

Record	Where it lands	Billed
A finished voiceover	A dataset row, ok: true	yes
The keyless demo	Key-value store, SAMPLE_AND_NOTICES, _demo: true	no
An item that failed	Key-value store, SAMPLE_AND_NOTICES, ok: false	no

So everything in the dataset is audio you can use. One failed item does not stop the batch: it is recorded in the store and the run moves on to the next item.

Worth knowing, because it is the thing that confuses people: if every item fails, the run still finishes green with an empty dataset. The reasons are all sitting in SAMPLE_AND_NOTICES in the key-value store.

The Overview table shows index, voice, format, durationSeconds, characters, audioUrl and textPreview. model and chunks are real fields but not among the columns, so open the row or download the JSON if you want them.

💡 What people use it for

• Narration for faceless video channels, where a week of scripts goes in as one texts batch.
• Audiobook and long-article reading, leaning on the automatic splitting so nothing has to be cut up by hand.
• Phone-menu prompts, where WAV is usually the format the phone system wants.
• Trying the same script in several voices before committing to one.

From a written post to a narrated slideshow, in three runs:

1. Run Reddit Text Cleaner on the post and copy the cleaned text it returns.
2. Paste that text here into Text / script and take the audioUrl from the row.
3. Put the audioUrl into Audio URL on Storyboard Video Generator, with your images in Image URLs, and you get an MP4 with your voiceover laid over it.

🚧 What it does not do

• No voice cloning, and no custom voices. Six voices, listed above, and that is the list.
• No SSML. There are no tags for pauses, emphasis or pronunciation. Punctuation and speed are the only controls over delivery.
• A rejected key stops the batch, rather than retrying its way down your whole list.
• durationSeconds can read 0 when the finished file could not be measured. The audio is still there, and the row is still charged.
• Joined opus and aac files are less proven than MP3. Long scripts in those two formats are stitched the same way, but MP3 is the one with playback evidence behind it. Use MP3 or WAV if a long script has to be right first time.
• No video, no captions, no music bed. This is the audio track only.
• It is not an official tool of any platform or company named on this page. Dami's Studio is independent and is not affiliated with or endorsed by any of them.

🧭 Which AI media actor do you need?

If you want	Use
A script read out loud as an audio file	This one
A finished vertical video with script, voice and captions	AI Faceless Video Generator
An existing video re-voiced in another language	AI Video Dubber
Images turned into a slideshow video, with your voiceover laid over it	Storyboard Video Generator
Scene images for a story before you voice it	AI Storyboard Generator
Word by word captions burned onto a video you already have	Auto Caption Burner
The other direction: speech in a video or audio file written out as text	Video & Audio Transcriber
A Reddit post cleaned up so a voice can read it	Reddit Text Cleaner

❓ Questions people ask

Do I need my own OpenAI key for text to speech?

Yes. There is no shared key behind this, and without one you get the demo record rather than audio.

Is my OpenAI key safe here?

It is a secret field, so it is not stored in the run's visible input. Use one you can revoke, as with anything you paste into a tool you did not write.

Can I use something other than OpenAI?

If it speaks the same /audio/speech shape, point baseUrl at it. Anything else will not work.

Why is my dataset empty when the run went green?

Every item failed. The reasons are in SAMPLE_AND_NOTICES in the run's key-value store.

Can I call it from code or connect it to an AI assistant?

Yes. The API tab has ready-made code for Python, JavaScript and the command line. For Claude, ChatGPT or another MCP client, connect https://mcp.apify.com/?tools=fetch-actor-details,dami_studio/ai-tts-voiceover. Either way the run happens on your Apify account at the same price, and the speech is billed to your OpenAI key as usual.

🆘 If something breaks

Open the Issues tab on the actor page. Include the run ID and which format and voice you used, and check SAMPLE_AND_NOTICES in the run's key-value store first, since the per-item reason is usually already written there.