The Two Limits: Characters and Licence
Text-to-speech is metered per character, which makes the free tiers easy to compare once you convert them into something meaningful. A rough and reliable conversion: 1,000 characters is about 150 words, which is about a minute of speech.
So a ten-minute YouTube script is roughly 1,500 words, or 10,000 characters. Hold that number — it is exactly one month of ElevenLabs' free plan, and one four-hundredth of Google's.
The second limit is the one that actually stops projects: whether you are allowed to publish the result. A free tier that forbids commercial use is fine for testing a voice and useless for a client video, a monetised channel or a product. These two limits are independent, and the tool with the most generous character allowance is often also the one with the cleanest licence — which is why the popular answer is usually the wrong one.
Before you pick anything, count your script. Our Word Counter gives the character total, and how to count characters online explains why spaces and markup count too — Google, for one, bills every space, newline and SSML tag in the input.
The Consumer Apps
ElevenLabs Best voices, strictest free plan
The quality leader, and the free plan is a demo rather than a working allowance. Its pricing page lists the Free tier as 10,000 credits per month — around 10 minutes of audio — with about 10 concurrent requests, access to text-to-speech, speech-to-text, sound effects, voice design and music, and commercial use not allowed.
Commercial licensing and instant voice cloning start at Starter, $6/month for 30,000 credits. Creator is $22/month for 121,000 credits, with a 50% first-month discount. For a monetised channel publishing weekly, $6 is the honest entry price — and it is a fair one, roughly 30 minutes of audio.
Use the free plan for what it is good at: auditioning voices before you commit. Do not build a workflow on it.
The Cloud Free Tiers Nobody Mentions
The big cloud providers sell TTS by the character and give away large monthly allowances permanently. They rarely appear in "best free AI voice generator" lists because they need an API call rather than a text box — which is a real cost, but a small one against the numbers.
Google Cloud Text-to-Speech Most characters, every month
Google's free tier is 4 million characters per month for Standard voices and 1 million per month for WaveNet voices, renewed monthly rather than a one-off trial. That is around 650,000 words of Standard audio a month — enough for a daily podcast, an entire audiobook backlog, or every page of a large site read aloud.
The counting is strict: the total input string is billed including spaces and newlines, and all SSML tags except <mark> count too. If your script arrives from a PDF or a CMS with stray characters and double spacing, AI Text Cleaner trims what you would otherwise be paying for.
Amazon Polly 5M standard characters
Polly's free tier is 5 million characters per month for Standard voices, with no expiry stated, plus time-limited allowances for the better voices: 1 million characters a month for Neural, 500,000 for Long-Form and 100,000 for Generative — each for the first 12 months only. After that: $4 per million characters for Standard, $16 Neural, $30 Generative, $100 Long-Form.
Read that expiry carefully. Build a workflow on Polly's Neural free tier and it becomes a bill on month thirteen — modest at $16 per million characters, but no longer free.
Azure AI Speech 0.5M neural characters
Azure's F0 free tier includes 0.5 million characters per month of neural text-to-speech, monthly and permanent. Smaller than the other two, but note what it is: half a million characters of *neural* voice, where Google's largest free bucket is Standard quality and Polly's neural allowance expires. Custom voice is not part of the free tier.
All three are pay-as-you-go services, so publishing what you generate is a normal commercial use — the opposite posture from a consumer free plan.
Local Models: Unlimited and Free
If you want no meter at all, run the model yourself. Two are worth knowing, and the licence difference between them matters.
Kokoro-82M is an open-weight model of 82 million parameters released under Apache 2.0, with 54 voices across 8 language configurations in v1.0. Its authors state plainly that it is Apache-licensed and has been deployed in commercial APIs, so commercial use is permitted. It installs with pip and runs locally, including on a free Colab instance. For a small model it is remarkably good, and commercial API providers reselling it charge around $0.65–$0.80 per million characters — which tells you what you are getting for nothing.
Piper is the other well-known local engine, and here is the detail most articles have not caught up with: the original MIT-licensed repository was archived on 6 October 2025, and development moved to a new home under the Open Home Foundation — where the project is GPL-3.0. It is still fast, still local, still free, and it powers Home Assistant and the NVDA screen reader. But GPL-3.0 is a copyleft licence, so if you are embedding a TTS engine in a product you distribute, that change is one to read before you ship. The Foundation is also openly looking for maintainers.
Local models have the same privacy advantage as local chat: the text never leaves your machine, which matters if you are narrating anything confidential. The same reasoning as our guide to free local AI chatbots applies here.
What No Free Tier Fixes
Three problems survive every plan on this page, and they cost more time than the character limits do.
Proper nouns and acronyms come out wrong. Product names, place names and initialisms are where synthetic speech reliably embarrasses itself, and no amount of free credits changes the default pronunciation. The fix is SSML — marking up a word with an explicit phonetic spelling — which every cloud service supports. Budget for it: on Google those tags count toward your character total, so a heavily marked-up script burns the allowance faster than its word count suggests.
Long scripts are not one request. Services cap the length of a single synthesis call, so an audiobook chapter becomes a loop that splits text, calls the API and stitches the audio back together. That is scripting work, not clicking work, and it is the real reason the generous cloud tiers stay missing from "best free tool" lists. ElevenLabs' free plan allows about ten concurrent requests, which is the same constraint seen from the other side.
Voice cloning is the part to be careful with. Instant voice cloning sits behind ElevenLabs' paid tiers, and Azure excludes custom voice from its free F0 tier entirely — so the free options mostly hand you stock voices, which is fine. When you do reach for cloning, the rule is simple and not primarily a licensing question: clone a voice only with the explicit permission of the person it belongs to. A stock voice you are allowed to publish beats a cloned one you are not.
Read the script aloud yourself before synthesising it, too. Sentences that work on the page often stall when spoken, and fixing the text is faster than fighting the voice.
Comparison Table
| Tool | Free Allowance | Commercial Use? | Watch Out For |
|---|---|---|---|
| ElevenLabs | 10,000 credits/mo (~10 min) | No, not on Free | Commercial licence starts at $6/mo Starter |
| Google Cloud TTS | 4M chars/mo Standard · 1M WaveNet | Yes — pay-as-you-go service | Spaces and SSML tags are billed; needs an API call |
| Amazon Polly | 5M chars/mo Standard; Neural 1M for 12 months | Yes | Neural, Long-Form and Generative free tiers expire at 12 months |
| Azure AI Speech | 0.5M neural chars/mo, permanent | Yes | Custom voice excluded from Free |
| Kokoro-82M | Unlimited — runs locally | Yes — Apache 2.0 | 54 voices, 8 language configs; you supply the hardware |
| Piper | Unlimited — runs locally | Yes, but GPL-3.0 copyleft | MIT repo archived Oct 2025; active fork is GPL-3.0 |
Which One Should You Use?
- You are publishing the audio and want the best free deal: Google Cloud TTS. Four million characters a month, renewed, and commercial use is simply how the service works.
- You want neural quality for free, permanently: Azure's 0.5M characters a month. Smaller bucket, better voices, no expiry date.
- You want the best-sounding voice and you are monetising: pay ElevenLabs $6. The free plan cannot be used for this, and no workaround changes that.
- You want zero meter and zero accounts: Kokoro-82M. Apache 2.0, commercial use allowed, 54 voices, runs on your own machine.
- You are embedding TTS in a product you distribute: Kokoro over Piper, purely on licensing — Apache 2.0 versus GPL-3.0 copyleft.
- You are narrating confidential material: local only. A script pasted into a web tool has left your machine.
- You are building a video: pair the voice with visuals from our roundup of free AI video generators, and check the script reads aloud well with Readability Checker first.