In short
On the query “ai voice generator”, 16 of the 19 organic results are product pages. They all advertise something free, and almost none of them says how many minutes of audio that buys you.
Quotas are counted in characters, in credits, in minutes, in hours, in videos or in numbers of voices. Only one unit lets you compare: the minute of audio, and the conversion factor is published by the vendors themselves, at a thousand characters per minute.
The real wall is not the counter, it is the commercial licence, missing from the free column at most vendors. And cloning a real person’s voice to publish it now runs into a property right in that voice in a growing number of states.
You type “ai voice generator”, you open the first five links, and you get five identical promises: free, realistic, no equipment. Not one of those five pages answers the only question that matters when you have a script to read out loud: how many minutes of audio can I produce before it stops, and am I allowed to use them?
We opened the pricing pages of the services that occupy this results page on 11 August 2026 and wrote down what they actually publish. It comes down to three findings: the units of measurement are never the same, the free quota converts into a common unit anyway, and it is the “commercial use” line that decides everything.
An AI voice generator is four different products
The keyword lumps together services that do not do the same job. Knowing which one you are after saves you from signing up in the wrong place.
Text to speech
You paste a text, you pick a voice from the catalogue, you get an audio file back. It is the most common product and the only one that is genuinely free somewhere.
For: video voice-over, e-learning module, phone greeting.
Voice cloning
You upload a recording and the service builds a synthetic voice that says whatever you want. Nearly always reserved for paid plans, and covered by law when the voice is not yours.
For: reusing your own voice without recording again.
Voice changer
You speak and the service swaps your timbre for another one while keeping your delivery. The result is livelier than a text read-out, but you have to act the take.
For: podcast, character work, anonymising.
Talking avatar
The voice comes bundled with a video and a synthetic presenter. The free quota is then counted in videos, not characters, and it is the tightest of the lot.
For: product walkthrough, internal message, training.
What Google serves on this query Of the 19 organic results measured on 11 August 2026, 16 are product pages, one is a Google Play listing, one is a YouTube video and one is a Reddit thread. A carousel of ten short videos sits at positions 11 to 20, and there is no “people also ask” block and not a single editorial article. In other words, Google has nothing to offer between the sales pitches. The top three are QuillBot, Canva and CapCut.
The only unit that lets you compare: the minute of audio
Voicebooking counts in characters. ElevenLabs and Fish Audio count in credits. TTSStudio counts in hours per month. Narakeet sells minutes. HeyGen hands out videos. Speechify counts nothing at all and advertises a number of voices. Six units for measuring exactly the same thing: the length of a sound file.
The good news is that the conversion does not have to be invented. The vendors publish it themselves, and you only have to read it.
ElevenLabs spells it out: “For V2 Multilingual models, 1 text character equals 1 credit.” Its grid gives the free plan 10,000 credits a month. And its own comparison table turns that quota into a duration.
Ten thousand characters, ten minutes. A thousand characters are therefore worth one minute of audio, and that ratio comes from the market leader on its own page. Two competitors publish the same kind of mapping and land in the same range: Fish Audio advertises 8,000 monthly credits and “Up to 7 minutes generation”, which works out at roughly 1,143 characters a minute; MiniMax Audio advertises 10,000 credits and “~12 mins of HD model”, roughly 833 characters a minute.
The conversion rule used in this article A thousand characters, spaces included, are worth one minute of audio. That is the figure ElevenLabs publishes. It slightly overstates the duration compared with Fish Audio and understates it compared with MiniMax: the spread between vendors runs from 833 to 1,143 characters a minute, about 15 % either side. For a quota calculation, that margin changes nothing.
One detail is worth flagging in passing, because it appears on no pricing page: the unit itself depends on the language. MiniMax says so in its FAQ.
Translated: depending on the language, the same text does not burn the same number of credits, and English is the cheapest case documented. If your script is in Spanish, German or Japanese, expect the counter to move faster than this article predicts.
Count your script before you sign up
Paste your text below. The counter gives you the number of characters, the matching audio duration, what that duration would cost at the per-minute rate Narakeet publishes, and above all how many times that script fits inside each free quota we measured. Everything is computed in your browser, nothing is sent anywhere.
Voice script counter
Characters, words, audio duration and how many generations each free plan allows.
The text stays in your browser. Spaces and punctuation count towards the vendors’ quotas, so they are counted here too.
1,079
189
1.1min
0.22€
At this length the plans capped at a thousand characters are already behind you. You need a plan that publishes at least 8,000 credits a month, otherwise you will generate your text in several pieces and stitch them back together.
How many times your script fits inside each free plan
| Service | What the free plan publishes | In characters a month | Your script fits |
|---|---|---|---|
| Voicebooking | 1,000 characters, 1 project, 3 downloads | 1,000 |
0 times |
| HeyGen | 3 one-minute videos a month | 3,000 |
2 times |
| Fish Audio | 8,000 credits, “Up to 7 minutes generation” | 8,000 |
7 times |
| ElevenLabs | 10,000 credits, “~10” minutes included | 10,000 |
9 times |
| LuvVoice | 10,000 characters a month, 3,000 per conversion | 10,000 |
9 times |
| MiniMax Audio | 10,000 credits, “~12 mins of HD model” | 10,000 |
9 times |
| TTSStudio | no zero-dollar plan on the grid | 0 |
0 times |
| ReadSpeaker | demo with no signup, no counter on display | no published cap | as many times as you like, but never for commercial use |
The credits of ElevenLabs, Fish Audio and MiniMax are counted here at one character per credit for text to speech, which is what each of them documents on its own page. The 3,000 characters for HeyGen are a conversion of its 3 one-minute videos, at a thousand characters a minute. Figures captured on 11 August 2026.
What the length of your script changes
| Length | What it means |
|---|---|
| Under a minute, below 1,000 characters | A script this short goes through anywhere, including the stingiest plan on the list. It is the format of a video intro, a phone greeting or an audio post. |
| 1 to 3 minutes, 1,000 to 2,999 characters | At this length the plans capped at a thousand characters are already behind you. You need a plan that publishes at least 8,000 credits a month, otherwise you will generate your text in several pieces and stitch them back together. |
| 3 to 10 minutes, 3,000 to 9,999 characters | A single generation eats most of a free monthly quota. Plan on finding the right voice first time, or pay for the month you produce in. |
| Over 10 minutes, 10,000 characters and up | No free plan in this table holds a script of that length in one month. At that volume, the per-minute rate Narakeet publishes becomes the right benchmark. |
Script over the quota? Nation AI does not speak the text, it writes it and trims it until it fits.
Hello and welcome. In two minutes I will show you how to prepare a voice-over script that fits inside a free quota. First point: write short sentences. One idea per sentence, a period at the end. Speech synthesis breathes where you put the punctuation, and nowhere else. Second point: spell out your numbers. Twenty-four reads better than the digits, and one thousand five hundred avoids a hesitant delivery. Do the same for the acronyms you want to hear letter by letter. Third point: mark your pauses with commas and periods, not with line breaks. The engines ignore layout, they read punctuation. Fourth point: read the text out loud before you generate it. You will hear the repetitions, the awkward transitions and the sentences that run too long, the ones that force the voice to break in the wrong place. Last point: count your characters before you paste. A free quota of ten thousand characters is about ten minutes of audio a month, and a single retake costs you as much as the first version. That is it. Now replace this text with your own and watch the counters move.
What each free plan actually gives you
Here is the full survey, service by service, with its position on the English results page. Anything in quotation marks is copied from the official pages as they render in English on 11 August 2026.
| Service (US rank) | Unit shown | What the free plan gives | Download | Commercial use when free |
|---|---|---|---|---|
| QuillBot (1) | none | “completely free of charge”, no figure anywhere | yes, button on the tool | not stated |
| Canva (2) | none | no figure on the tool page | not measured | not measured |
| ElevenLabs (4) | credits | 10,000 credits, “~10” minutes | yes | no, listed on the $6 Starter plan |
| Voicebooking (6) | characters | 1,000 characters, 1 project | 3 downloads | yes, but the wording only exists on the French page |
| MiniMax Audio (10) | credits | 10,000 credits, “~12 mins of HD model” | yes | yes, “for songs created during free period” |
| Narakeet (13) | minutes bought | no zero-cost plan, 30 minutes for €6 | yes | not applicable |
| ReadSpeaker (outside the top 20) | none | demo with no signup, “unlimited voice samples” | yes | “strictly forbidden” |
| LuvVoice (outside the top 20) | characters | 10,000 characters a month, 3,000 per conversion | MP3, files kept 72 hours | no, listed on the $7.75 Plus plan |
| Fish Audio (outside the top 20) | credits | 8,000 credits, 7 minutes, 500 characters per generation | yes | no, struck through on the free tier |
| TTSStudio (outside the top 20) | hours a month | no zero-dollar plan, entry tier $9.99 for 10 h | unlimited on paid plans | not applicable |
| Speechify (outside the top 20) | number of voices | “10 robotic sounding voices”, natural voices at $29 a month | not stated | natural voices on the paid plan |
| HeyGen (outside the top 20) | videos | 3 one-minute videos a month | with a watermark | watermark removed on the $29 plan |
Voicebooking, the tightest quota and the widest licence
The sixth result gives you a thousand characters, so one minute of audio, a single project and three downloads. The last two lines carry a red cross on its own grid: they are limitations, not features.
In exchange, it is the only one in the panel that writes the exact opposite of ReadSpeaker on rights, and it writes it in one language only. The French version of the same page carries a FAQ entry asking whether the generator comes with limited usage rights, and answers that it does not: you may use it on any means of communication, «au niveau mondial, à perpétuité», worldwide and in perpetuity. That FAQ block does not exist on the English page, which stops at the grid. One minute of audio, but a minute you are allowed to publish.
ReadSpeaker, unlimited and forbidden
ReadSpeaker offers a demo with no signup. Its help page confirms it: “Yes, our text to speech demo is completely free with no signup required. You can test unlimited voice samples, try different languages, and download audio files.” No counter, no credits, no form. And, two blocks further down the same page, the line that cancels all of it.
Two services in the same panel therefore say the exact opposite of each other. One gives you a minute you can use anywhere, the other gives you unlimited time you can use nowhere, the moment money is involved.
ElevenLabs, the benchmark, but not for publishing
Ten thousand credits, ten minutes, 128 kbps audio. The ElevenLabs free plan is comfortable for auditioning voices. The “Commercial License” line only shows up from the Starter plan at $6 a month: it sits in the “Everything in free, plus” block, which is a roundabout way of saying it is missing below. Instant voice cloning lives in the same place.
Fish Audio, seven minutes under a promise of unlimited
The Fish Audio pricing page opens with “Unleash your creativity with unlimited generations”. The free tier, right underneath, shows 8,000 monthly credits and “Up to 7 minutes generation”, with a cap of 500 characters per generation. Two lines in that same block are struck through: enhanced voice cloning and commercial use. Its FAQ is blunter still: “Free plan users can only use generated content for personal, non-commercial projects.”
The per-generation cap deserves a word. With 500 characters maximum in one go, a two-minute script has to be cut into five pieces, generated separately, then reassembled in an editor. The monthly quota is not the only lock: the size of a single request is another one, and the grids file it under small print.
LuvVoice, the most generous on the file itself
Ten thousand characters a month, three thousand per conversion, more than 200 voices, MP3 download and a custom voice with 500 dedicated credits. On features alone it is the most complete free plan in the panel. The “Unlimited commercial rights” line only appears from the Plus plan at $7.75 a month, and the free card states that files are kept for 72 hours.
Speechify, a free generator with no natural voice
The Speechify page that ranks on this query is headed “FREE AI VOICE GENERATOR”. Its pricing grid describes the free plan in two lines flagged with an orange exclamation mark: “10 robotic sounding voices” and “Text to speech features only”. The “1000+ high quality, natural voices” sit in the Premium column, at $29 a month.
It is the widest gap in the panel between the headline of the product page and the grid published by the same vendor. The service is genuinely free, just not for what the query is looking for: a voice that does not sound like a machine.
HeyGen, three minutes a month and a watermark
HeyGen does not sell voice, it sells video with a synthetic presenter. Its free plan is therefore counted in videos: “3 videos per month”, “Videos up to 1 min”. Three minutes a month is the tightest quota in the panel after Voicebooking. Two interesting lines sit in the paid column at $29: “Watermark removal” and “Voice Cloning”. Worth noting for anyone comparing across borders: the same Creator plan is priced at 25 € on the French and Spanish versions of the page, so this is a different grid rather than a conversion.
The ones with no zero-cost plan at all
Two services in this survey publish no free offer whatsoever, and that is useful information in itself: it saves you a pointless signup.
TTSStudio, which does not make the English top 20 but ranks in France and Spain, lines up three paid plans counted in hours of generation a month: 10 h for $9.99, 50 h for $29.99, 170 h for $99.99. No zero-dollar column. The page shows a one-hour countdown under each plan.
Narakeet, thirteenth, does not sell a subscription at all: “The plans are one-time payments to purchase additional capacity, based on the duration of audio/video materials you produce or transcribe, not recurring subscriptions.” Its grid is the most readable of the lot, because it states the price of a minute outright: €0.20 for 30 minutes, €0.15 for 300, €0.10 for 1,000, €0.08 for 2,500, €0.05 for 10,000. Note that Narakeet publishes its grid in euros, including on its English pages.
It is also the only grid in the panel that lets you put a price on a free quota. The ten minutes ElevenLabs gives away are worth €2 at Narakeet’s entry rate. The most generous free month on the voice market is worth roughly the price of a coffee.
The ones that publish no figure at all
QuillBot takes the first position with a page that promises “Use completely free of charge, with zero hidden fees and no sign-up required”, offers a Download button, and never states a limit. Its own FAQ, asked whether the tool is free, answers that the generator “does not cost money to use” and leaves it there. The interface counts your words as you type, which suggests there is a cap somewhere, but nothing on the page says what it is.
Canva, second, does the same with a very polished tool page that explains the process in three steps and never mentions a quota, a credit or a maximum duration: its voice generator draws on the account’s shared AI allowance, whose figure lives elsewhere, on the subscription pages. CapCut, third, goes further still: the word “free” comes back twenty-five times on its page, FAQ included, and no number ever follows it. Adobe Firefly could not be measured at all: the site refuses automated rendering, and we would rather write “not measured” than copy a figure found somewhere else.
The wall is not the counter, it is the licence
Go back to the last column of the big table. Of the eight services that publish a usable free offer, only three allow commercial use without paying: Voicebooking in its FAQ, MiniMax for tracks created during the free period and, with an important caveat, HeyGen provided you accept the watermark. At ElevenLabs, LuvVoice and Fish Audio, the commercial licence is a paid line. At ReadSpeaker it is explicitly forbidden.
A monetised channel is commercial use The question comes up constantly and the answer is nearly always yes. A YouTube video with ads, a sales funnel, a company post, a paid course: all of that falls under “commercial use” as these services define it. Generating the voice-over of a monetised video on a plan that forbids it puts you in breach with the vendor, whatever your view count.
Three more locks hide in the same grids, and none of them is ever advertised.
- The per-generation cap. Fish Audio limits each request to 500 characters, LuvVoice to 3,000 on the free plan. A long script has to be cut up and glued back together.
- The retention window. LuvVoice keeps files for 72 hours. If you do not download straight away you have to regenerate, and regenerating burns the quota a second time.
- The number of downloads. Voicebooking grants three, which is not the same as three generations: you can listen as much as you like, you only leave with three files.
Cloning someone’s voice: what US law says
Among Google’s own completions for this query sits “ai voice generator celebrity”. It is one of the most frequent searches in the cluster, and not one page in the top 20 explains what you are risking. There is no single federal statute on this, but two 2024 texts changed the picture and both are short.
The first is a state law. Tennessee’s Public Chapter 588, better known as the ELVIS Act, renamed the Personal Rights Protection Act of 1984 the “Ensuring Likeness, Voice, and Image Security Act of 2024” and added a definition that leaves no room for argument:
“Voice” means a sound in a medium that is readily identifiable and attributable to a particular individual, regardless of whether the sound contains the actual voice or a simulation of the voice of the individual
The same act rewrites the core rule as “Every individual has a property right in the use of that individual’s name, photograph, voice, or likeness in any medium in any manner”, and creates civil liability for anyone who “publishes, performs, distributes, transmits, or otherwise makes available to the public an individual’s voice or likeness” without authorisation. It took effect on 1 July 2024.
Three practical consequences, as general information and not as legal advice.
- Consent is the first condition. Cloning your own voice raises none of this. Cloning a relative’s, a colleague’s, a performer’s or a politician’s does, even with no intent to harm.
- The right survives the person and covers the tool. The Tennessee text also makes it actionable to distribute software whose primary purpose is producing a particular identifiable individual’s voice without authorisation.
- Where you are matters. Right of publicity is state law and it varies: California, New York and Tennessee are the strictest, some states have no statute at all. The rule of thumb that travels is simple: no consent, no publication.
The second text is federal, and narrower. On 8 February 2024 the FCC adopted a Declaratory Ruling recognising that calls made with AI-generated voices are “artificial” under the Telephone Consumer Protection Act, which, in its own words, “makes voice cloning technology used in common robocall scams targeting consumers illegal”. If you distribute in the European Union, add Article 50 of Regulation (EU) 2024/1689, applicable since 2 August 2026, which requires providers of systems generating synthetic audio to mark their output in a machine-readable format and requires whoever publishes a deep fake to disclose that it was artificially generated.
Marking is the vendor’s job, disclosure is yours The technical watermark is on the company that builds the generator. Saying so is on you. In practice you do not have to produce the machine-readable signal, but you do have to write somewhere that the voice in your video is synthetic as soon as it imitates a real person. One line in the description is enough, and it is the cheapest evidence of good faith you will ever produce.
The subject has a darker side, the cloned voice used to deceive. We covered that separately in our investigation into the AI voice phone scam, where we show how many seconds of recording it really takes to clone a voice.
Getting a voice that does not sound like a robot
Once the service is picked, the quality of the result depends less on the engine than on the text you feed it. The five settings below cost no credits at all and change everything.
- Write short sentences. Speech synthesis breathes at the punctuation and nowhere else. A forty-word sentence with no comma will be read in one go, without a breath.
- Spell out your numbers. “One thousand five hundred” is always pronounced correctly, “1,500” depends on the engine. Same for acronyms you want spelled out and for units.
- Do not rely on layout. Line breaks and bullets are ignored. A pause is requested with a period, a colon or a comma.
- Choose the voice before writing the long version. Try three voices on two sentences, not on the whole script: it is the best way not to burn a monthly quota on auditions.
- Read it out loud before generating. You will hear the repetitions, the awkward joins and the sentences that run too long. What you cannot say yourself, the machine will not say better.
Before
Our $29.90/mo plan includes 3 modules (CRM, support, HR) available 24/7 from any device, with no commitment and a 14-day trial.
After
Our plan costs twenty-nine dollars and ninety cents a month. It includes three modules: customer relations, support, and human resources. Everything is available at any time, from any device. No commitment, and a fourteen-day trial.
The version on the right has more characters, so it costs more credits. That is the real trade-off of this market: a text that reads well out loud is a longer text, and a longer text costs more. The right strategy is not to write short, it is to write dense and then cut everything that cannot be heard.
The mistakes that cost you a quota
- Generating the whole script to audition a voice. Two sentences are enough to judge a timbre. You test a voice catalogue on a sample, not on the production.
- Forgetting to download. At LuvVoice the file disappears after 72 hours and the second generation goes through the till again.
- Pasting text with markup. Markdown asterisks, bullets and headings sometimes end up read out loud. Clean before you paste.
- Taking the homepage promise for the grid. “Unlimited”, “free” and “no signup” are landing-page arguments. The truth is in the left-hand column of the pricing table.
- Confusing free with royalty-free. Those are two separate questions, and the second one decides what you will be allowed to publish.
Nation AI does not record the voice, it writes the text the voice will read
Let us be straight about it: Nation AI does not produce audio files. It is not a voice generator. What the counter above shows, though, is where the game is really played on this subject: the quota is spent in characters. A badly calibrated script means a blown quota, a retake, and a second trip to the till.
That is exactly the job you can hand it: writing a voice-over script calibrated to a duration, trimming it to an exact character count, replacing digits with words, breaking up sentences that run too long. For plain rewriting of an existing text, our rephrasing tool goes straight to the point.
A script that fits the quota
Nation AI writes, trims and cleans the text your voice generator is going to read. Free to try, no signup.
Frequently asked questions
What is the best free AI voice generator?
It depends what you plan to do with the file. For commercial publishing without paying, Voicebooking is the only one in the panel that explicitly allows it, but it only gives one minute. For trial volume, ElevenLabs and LuvVoice give ten minutes a month, with no commercial licence. To listen without limits and publish nothing, the ReadSpeaker demo has no counter at all.
Can you generate an AI voice without signing up?
Yes, but rarely with a download. ReadSpeaker offers a demo that is “completely free with no signup required” and lets you download audio files, with all commercial use excluded. QuillBot also promises “no sign-up required” and offers a download button, without publishing any limit. Elsewhere, the no-account trial stops at the moment you try to export.
How many characters make one minute of audio?
About a thousand. That is the ratio ElevenLabs publishes, whose free plan pairs 10,000 credits with “~10” included minutes, given that one character equals one credit there. Fish Audio is a little more generous, at roughly 1,143 characters a minute, MiniMax a little tighter, at roughly 833.
How do you get the voice as an MP3?
Check the download line of the grid before signing up, because that is where most free offers stop. LuvVoice advertises MP3 from its free plan, Voicebooking caps you at three downloads, HeyGen exports with a watermark until you move to a paid plan.
Can you make a celebrity voice say whatever you want?
No. In Tennessee, the ELVIS Act of 2024 gives every individual a property right in their voice, explicitly including a simulation of it, and makes publishing an unauthorised voice actionable. Other states protect the same thing through their right of publicity, with different wording and different limits. This is general information, not legal advice.
Can a free AI voice be used on a monetised YouTube channel?
A monetised channel is commercial use. Of the eight free offers measured, only three allow it, and ReadSpeaker forbids it in writing. Check the “commercial use” line of the free column before you publish, not after.
Do you have to disclose that a voice was AI-generated?
As soon as the voice imitates a real person, yes, and for two reasons. Disclosure is what separates a synthetic performance from an attempt to pass it off as authentic, which is what right-of-publicity claims turn on. And if you distribute in the European Union, Article 50 of the EU AI Act, applicable since 2 August 2026, requires whoever publishes a deep fake to disclose that the content was artificially generated.
Why do non-English voices sound worse?
Voice catalogues are larger in English and multilingual models are often trained on English first. Two settings make up a lot of ground: spelling out numbers and cutting long sentences. Try several voices on the same sentence too, because the gap between two voices from the same vendor is wider than the gap between vendors.
What to remember
The AI voice market sells free without ever saying how much. Convert everything into the same unit, the minute of audio, and the ranking sorts itself out: ten minutes a month at the leader, seven at its direct rival, one at the tightest of the panel, zero at two services that rank in Europe. Then look at the line below, the licence line, because that is the one that decides whether the file is any use to you.
And keep in mind that a person’s voice is not raw material. Since 2024 it is named property in Tennessee law and a cloned voice in a robocall is illegal under FCC rules. Saying that it is AI is usually enough to stay on the right side.
For the rest of the production chain, we measured the real price of a second of AI video the same way, along with the quotas and rights of AI music generators and what actually blocks you on song generators with lyrics. For the editing that comes after the voice-over, see our pick of free AI video editing tools.
Sources. Pricing and product pages consulted on 11 August 2026: QuillBot, Voicebooking, ElevenLabs, Canva, ReadSpeaker, LuvVoice, MiniMax Audio, Fish Audio, TTSStudio, Speechify, HeyGen, Narakeet, CapCut. Tennessee Public Chapter 588 (House Bill 2091), Ensuring Likeness, Voice, and Image Security Act of 2024, effective 1 July 2024. FCC Declaratory Ruling of 8 February 2024, news release DOC-400393A1. Regulation (EU) 2024/1689, Articles 50 and 113, on EUR-Lex. Search results positions measured on 11 August 2026.
