The first wall is arithmetic. A new ElevenLabs account gets 10,000 credits a month on Free, and on the Multilingual v2 model one character costs one credit — so that is roughly ten minutes of narration, about 1,500 spoken English words. It covers one short video. The second wall is licensing: Free carries no Commercial License, which appears only at Starter, $6/month for 30,000 credits (elevenlabs.io/pricing).
As of early October 2026 the company is still competing on quality, not generosity: it launched its Eleven v4 model with a promotional credit bump on Creator plans and above through October 12.
If you only need text spoken aloud, or a file you own, there are three routes out, ordered by effort. All three are free.
Official image from the ElevenLabs website.
Step 1: Have the browser or the OS read it to you
Edge’s Read Aloud, the Spoken Content feature on iOS and macOS, and a long list of browser extensions will speak almost any page, PDF or document out loud. There is no account, no word counter, and they work offline.
What it will not do is hand you an audio file. The sound stays on the device, so you cannot drop it into a video, and the output carries no commercial rights — fine for reading to yourself, not for invoicing a client.
The quality gap is real but narrower than it once was. Classic system voices sound plainly synthetic; Edge’s online neural voices are better, though still flatter than a paid engine. Rule of thumb: if the listener only needs to understand, this is enough; if the listener is paying, move on.
Step 2: Spend a free tier whose units are friendlier
Every vendor counts differently.
ElevenLabs Free gives 10,000 credits a month — about ten minutes, about 1,500 words. Microsoft Azure Speech’s free F0 tier gives 500,000 neural characters a month at no charge; at roughly five characters per word that is close to 100,000 words of narration, an audiobook-sized allowance that resets monthly, and Azure permits commercial use under its terms. Cartesia and Fish Audio sit in between with 20,000 and 8,000 monthly credits respectively (check official sites; these change).
The catch with Azure is the surface, not the quota: it is a developer console, so you paste text and call an API rather than drag clips on a timeline. The catch with the smaller free tiers is granularity — a per-request cap, a public cloning slot, and no rollover, so unused credits vanish at the reset.
Three questions worth answering before you commit
Is a free tier ever as good as ElevenLabs? For flat narration, close enough that most listeners will not notice; for emotional delivery and cloning from a short sample, no — that is what paid tiers sell.
How long does 10,000 credits actually last? One pass over a 1,500-word script, with no room for retakes. Budget it like a sample, not a session.
Do I have to give a card? Not for ElevenLabs Free. Azure’s free tier still needs an account with billing attached, even when usage stays free.
Step 3: Run the model yourself
When the quotas run out, the ceiling disappears if you host the model. Two open projects cover nearly every case.
Kokoro is an 82-million-parameter model with Apache-licensed weights, eight languages and 54 voices at v1.0, and it runs on ordinary hardware; a thousand characters of input is about a minute of output, and rented APIs serve it for under a dollar per million characters. Piper takes the other trade-off: pip install piper-tts, GPL-licensed, embedding espeak-ng for pronunciation, fast enough to run in real time on a laptop and already used by Home Assistant and NVDA. It is the more robotic of the two; Kokoro is the one that gets closest to commercial quality on neutral narration.
Neither clones a specific voice from a ten-second sample the way ElevenLabs’ paid tiers do, and setup costs an afternoon of Python. But once it runs, there is no monthly bill and no per-character meter.
Where free voice work usually goes wrong
Treating credits as minutes. A 10,000-credit month is ten minutes, not ten thousand. Convert to minutes before planning, not after.
Publishing unlicensed audio. Free tiers that withhold commercial rights are the norm. Generate your sample on a free tier; register for the paid plan before the first upload that earns money.
Rendering in one take. One mispronounced name burns the whole generation. Split the script into paragraphs so only the bad line is redone.
What to do next
Match the route to the job and stop there. Listening, proofreading or accessibility: system voices are already installed and free forever. A one-off file under about 1,500 words: ElevenLabs Free, or Azure’s 500,000 characters if the output needs a license. Anything longer, recurring or private: install Kokoro and treat the setup afternoon as the price of never paying again.
For the paid tiers themselves, see the ElevenLabs tool page and the ElevenLabs alternatives guide; the free AI voice generator roundup covers the license fine print.
Compiled from public pricing and product pages (elevenlabs.io, Microsoft Azure Speech, Hugging Face, GitHub) in early October 2026. Prices and free tiers change; check official sites. This guide is independent and accepts no payment for placement.