What per character billing means in text-to-speech

What does per-character billing actually mean in text-to-speech?

The service counts the characters in the text you send and multiplies by a rate, quoted per thousand or per million. It bills the text going in, not the audio coming out, so the price is known before any sound exists. Every request is counted on its own, which is why a second take costs as much as the first.

Per character billing is the default meter for cloud text-to-speech: Google Cloud, Amazon Polly, Microsoft Azure and ElevenLabs all use it. Every pricing page shows the rate. This guide covers what they leave out: what gets counted, how the count becomes minutes, and when the same sentence is counted twice.

The bill is set before the audio exists

A character count is fixed the moment you press the button. The length of the audio is not: it depends on the voice, the pace and the pauses the model picks. Billing the input lets a provider quote a price up front. It also means the meter cannot tell a first draft from a fifth attempt at a brand name. Both are text arriving, to be counted.

Billable characters: what each service counts

Google, Azure and ElevenLabs say outright that spaces and punctuation count. They differ on markup. Google bills every SSML tag except <mark>, Azure bills all markup except <speak> and <voice>, ElevenLabs counts SSML and audio tags, and Amazon Polly does not bill SSML tags at all. Azure also counts each Chinese character twice.

Amazon's request limit shows its rule at work: one synchronous request may hold 6,000 characters in total, of which no more than 3,000 can be billed, because the tags are free.

Spaces are about a fifth of the bill

Whitespace is easy to forget because nobody reads it. The DAISY Consortium puts an average novel at 90,000 words and 423,000 characters. Roughly one space separates each word, so about one character in five on that bill is a space. You cannot remove them without breaking the text, but you can stop paying for the extras: double spaces, blank lines and indentation copied in from a word processor are all billed at the full rate.

How many characters is one minute of audio?

Between about 700 and 1,000. ElevenLabs maps 1,000 characters to about a minute on its own price list. A Microsoft forum answer works it out at about 750, from 150 words a minute and five characters a word. DAISY's figure of 4.7 characters per word gives about 705 at the same pace.

The spread comes from reading speed and word length. ElevenLabs' 1,000 implies a brisk read, about 213 words a minute at 4.7 characters a word. Its API help pages use the same ratio: a 10,000-character request is about 10 minutes of audio.

Why we use 1,000 for budgeting

The higher figure is the cautious one for working out what a plan covers: a slower narrator gets more minutes from the same characters, not fewer. Our cost guides use 1,000 and say so. To find your own figure, render one paragraph and divide its character count by its length in minutes. The arithmetic for a full piece of writing is in what it costs to narrate a 5,000-word article.

ElevenLabs credits are characters with a new name

ElevenLabs says credits were previously called characters and the value is unchanged. On its v2 Multilingual models one character is one credit; Flash and Turbo cost 0.5 to 1 credit through the API. The Creator plan's 121,000 credits are shown as about 121 minutes, so a thousand credits buys roughly a minute.

The rename came because speech to text, music, sound effects and dubbing now draw on the same monthly pool. For text-to-speech the old rule stands: every letter, number, punctuation mark and space counts, and so do SSML and audio tags.

Turning credits into minutes

The formula is the same for any plan: credits, divided by credits per character for your model, divided by characters per minute.

• Free: 10,000 credits, about 10 minutes.
• Creator, $22 a month: 121,000 credits, about 121 minutes, with an extra minute at about $0.18.
• The API: billed in dollars, not credits, at $0.08 per 1,000 characters on the multilingual models and $0.04 on Flash and Turbo. The Creator plan's API listing includes 275,000 characters of v4 speech, shown as about 275 minutes.

The website and the API do not use the same unit, so check which one your plan describes before you divide. Unused plan credits roll over for up to two months while you stay subscribed, and are lost if you cancel or downgrade.

Regenerating a line is billed as new text

Yes, with narrow exceptions. ElevenLabs deducts credits every time you press Generate, including after a text edit. Its website gives two free regenerations only if the text, voice and model are unchanged, within two hours, without leaving the page, and never through the API. The developer APIs simply count each request.

The free regeneration covers keeping the words and nudging the voice settings. It does not cover the common case: you hear the line, change a word, and generate again. On v3, each Generate returns two takes and charges for one. On Google, Amazon and Azure a request is simply a request.

What a re-take habit does to an allowance

Say you render every line three times before you keep one, which is ordinary for narration. The 121 minutes in a Creator plan then become about 40 minutes of finished audio. Two takes gives about 60. Nothing about the rate changed; the count simply included the takes you threw away. DAISY found the same effect at book scale: titles that needed multiple corrections could end up two or three times the price of a single conversion. The re-takes, failed renders and overage behind that multiplier get their own guide, the hidden costs of cloud TTS.

Per character versus per second of audio

Some newer models bill the output instead. Google's Gemini text-to-speech charges audio tokens at 25 per second of speech, so a slow read with long pauses costs more than a brisk one. Per-character billing has the opposite property: the same script costs the same however slowly it is read.

At Gemini 2.5 Flash TTS's rate of $10 per million audio tokens, a minute of speech is 1,500 tokens, or about $0.015, plus a small charge for the text going in. Neither model is cheaper in general. Per character suits a calm, pause-heavy read; per second suits fast copy. Both bill a re-take in full. The three shapes of pricing, side by side, are compared in text-to-speech pricing models compared.

What changes when nothing is counted

A model that runs on your own Mac has no meter, so the count stops mattering. txt2mp3 is a one-time purchase; after it, a re-take, a pronunciation fix or a full re-render of a chapter costs nothing. A cloud plan still wins for a few minutes a month, on Windows, or for a large voice catalogue.

txt2mp3 runs VoxCPM2 on Apple silicon. The price is one payment, $29 for the launch batch, rising to $69 and then $99 as copies sell. There is no character count in the app except one: the trial speaks up to 120 characters at a time, enough to judge a voice on your own sentence. After activation the meter is gone, and so is the reason to stop at the first take that is good enough. As a side effect, the script never leaves the Mac.

If you narrate ten minutes a month, ElevenLabs' free plan costs less than any purchase, and speech on a server, a phone or Windows needs the cloud. What an hour costs on each service is worked out in our guide to the text-to-speech cost per character.

Frequently asked questions

Where the numbers came from