Voice a 5,000-Character Script and Deliver the MP3: ElevenLabs in the Browser, OpenAI in One Command, and the Licence Line

A client sends a 5,000-character script, about seven hundred words, and wants it read by a clean voice as an MP3 they can lay under a video. This page makes that file two ways, in the ElevenLabs browser app and with one request to OpenAI's speech endpoint, then walks the listen-through, the export and the two lines of terms that decide whether you may sell what you made. It reads the two providers' own guides for the steps and their own pages for the money.

What you hand over

One MP3 (or WAV, if the client's editor asks) of the whole script in one voice, with the names pronounced right, at the pace the client asked for, plus a one-line delivery note: which plan or model produced it and, if OpenAI did, that it must be disclosed as AI-generated to the audience. That last line is not politeness; it is in the provider's usage policies.

Step one: the script, as the model reads it

ElevenLabs' guide is blunt about input: the model reads what is written. A misspelled word is pronounced as spelled. Digits are a trap, because the number 1 is written identically in dozens of languages and pronounced differently in each, so the guide says to write numbers out as words, especially on multilingual models; the web app runs a normalisation step that spells out numbers and symbols for you, but the API does not. Punctuation shapes the delivery, so fix it. Emojis and unusual symbols make the output unstable, so remove them.

Then count characters, because every price on this page and one hard limit are per character. Eleven v3 takes 5,000 characters in one generation, which is exactly this script; Multilingual v2 takes 10,000; Flash v2.5 takes 40,000. A longer script is several generations, joined in your editor.

Path A: in the browser with ElevenLabs Text to Speech

Open the Text to Speech page and paste the script into the input box. Now choose the voice, from your Voices at the bottom left, before you touch anything else; the guide's order of importance is voice, then model, then settings, and it calls the voice the decision with the biggest effect on the output. Your Voices hold the curated Default voices, anything you saved from the Voice Library (reached through Voices, then Explore, with filters for language, accent, use case and quality), and your own clones. For a language other than English, pick a voice trained in that language, or the read will carry the wrong accent.

Choose the model second. If the script is English, ElevenLabs recommends an English-only model as more stable. Multilingual v2 is the most stable on long-form generations; Eleven v3 is the most expressive and takes audio tags for delivery; Flash v2.5 is the fast, cheap one, at 50% lower price per character on the API. Leave the sliders at the guide's common settings, stability around 50, similarity around 75, style at 0, speed at 1.0, and click Generate.

Path B: one command to OpenAI's speech endpoint

OpenAI's speech endpoint takes three things, the model, the text and the voice, and its command-line client does the whole job:

openai audio:speech create --model gpt-4o-mini-tts --voice coral --instructions "Speak in a cheerful and positive tone." --input "Today is a wonderful day to build something people love!" --output speech.mp3

Put the script in --input and the client's direction in --instructions; the guide says gpt-4o-mini-tts takes instructions for accent, emotional range, intonation, speed, tone and whispering. There are 13 built-in voices and the guide recommends marin or cedar for quality; the older tts-1 and tts-1-hd models take a smaller set of nine. The file is MP3 by default, and response_format can ask for opus, aac, flac, wav or pcm if the editor wants uncompressed audio.

The listen-through

Generation is not deterministic. The ElevenLabs guide says the same text and the same sliders produce a different take each time, and the sliders set how wide that range is rather than fixing a result. So the working method is to listen once through, note the lines that stumble, and regenerate only those. For a lively, dramatic read, set stability lower and generate a few times until one lands; for a serious read, set it higher and you will need fewer takes. Style exaggeration, the guide says, is best left at 0. Where a name or an abbreviation comes out wrong, the fix is in the script (a phonetic respelling, a written-out word), not in a slider.

Export the file the client asked for, and keep a copy on your own disk, for a reason under the terms below.

What a script costs

The chart above prices this exact script at each published per-character rate. On the API, ElevenLabs charges $0.10 per 1,000 characters on v2 and v3 voices and $0.05 on Flash and Turbo, in US dollars rather than credits, so the script is $0.50 or $0.25. OpenAI charges $15.00 per 1M characters on tts-1 and $30.00 on tts-1-hd, so $0.075 or $0.15; gpt-4o-mini-tts is priced per audio token, $12.00 per 1M output tokens, and the page prints no per-character figure for it, so it is not on the chart.

Inside the ElevenLabs app the unit is the credit, and Text to Speech costs 1 credit per character, so this script is 5,000 credits. The Free plan's 10,000 credits a month voice two scripts of this size. Starter, $6 for 30,000 credits, voices six. Creator, $22 for 121,000, voices twenty-four; Pro, $99 for 600,000, voices a hundred and twenty. Credits are shared across every ElevenLabs product and unused ones roll over for up to two months. The AI job cost calculator runs the API rates for any script length.

The plan decides whether you may sell it

ElevenLabs' help centre says the free plan does not include a commercial license and cannot be used for any commercial purpose, and that free-plan output must credit elevenlabs.io in the title. Every paid plan includes a commercial license outside Beta Services. So the two free scripts a month are for your own work, and Starter at $6 is the first door through which a client's voiceover can walk. The API is billed in dollars outside the plans, but the licence question is the same: it is the account's plan that answers it.

Whose voice it is, and for how long

You may use generated content commercially as long as you hold the necessary intellectual property rights to it. That means rights to the script, and rights to any voice you cloned: a Professional Voice Clone stays in your account only unless you share it. A Voice Library voice is someone else's clone, shared to a marketplace, and the Voice Library guide describes how it can leave: a voice with a notice period stays available to you for up to 2 years after its owner removes it, with email and in-app notice; a voice without one becomes unavailable immediately. If a client's whole channel is built on a library voice, filter for a notice period before you choose it. Library voices are also not available through the API on the free tier.

What survives cancellation

ElevenLabs says you keep a commercial licence forever to whatever you generated during a paid subscription. The files are a different matter: under the free tier they remain accessible only for an unspecified amount of time with no guarantee. Download before you downgrade. On the OpenAI side the file is yours from the start, since the services agreement assigns the Output to the customer, and the terms that bite are the account ones: ElevenLabs lists conduct that ends an account, such as fraudulent transactions or submitting misleading or unauthorized content, and allows one free account per user and IP.

If the voiceover is sold through a marketplace

A voiceover gig on Fiverr pays after a flat 20% commission and a 14-day clearing period; withdrawal to PayPal has a $1 minimum, no fee, and lands within 24 hours. Against a generation cost under a dollar, the marketplace's 20% and its two-week hold are the numbers that move your price, and the first freelance job page reads Fiverr's rules in full.

What to do today

Take a 5,000-character script you already have. Rewrite the numbers as words and check the names. Voice it once in ElevenLabs with the voice chosen first, once through OpenAI's endpoint with an instructions line, and listen to both against the client's brief. Then read the plan line on your ElevenLabs account before you send either file anywhere, because that line, not the price, decides whether the job is sellable.