Kokoro TTS: What It Is and How to Use It Free
Kokoro is an 82-million-parameter open-weight text to speech model that runs on modest hardware. Here's what it does well, where it falls short, and how to try it.
Kokoro is an open-weight text-to-speech model with 82 million parameters, small enough to run on a laptop CPU yet close enough to commercial quality that most listeners cannot tell the difference in a short clip. It is released under an Apache licence, which means you can use its output commercially without paying anyone.
That combination — genuinely small, genuinely good, genuinely free — is why it turned up everywhere within months of release, and why a free TTS tier built on it is a real product rather than a crippled demo.
Why 82 million parameters is the interesting number
Model size is usually reported like a score, where bigger wins. Kokoro is interesting precisely because it is not.
For comparison, commercial speech models are typically an order of magnitude or more larger and run only in a vendor's data centre. Kokoro fits in a few hundred megabytes. It runs on consumer hardware without a GPU. Inference is fast enough that generating a minute of speech takes seconds rather than minutes.
That changes the economics rather than the output. Running Kokoro costs roughly $0.007 per 1,000 characters, against about $0.10 for a top-tier commercial engine — a spread of roughly 14×. The gap does not buy basic intelligibility, which Kokoro already has. It buys emotional range, voice cloning and consistency across very long passages.
For narration, explainers, study material and most video voiceover, that gap is not worth paying for. For an audiobook with emotional arc or an advert where the read is the product, it is.
What Kokoro is good at
Clear, neutral narration. This is the model's centre of gravity. Explainer videos, documentation read-alouds, study notes and faceless-channel scripts all sit comfortably inside what it does well.
Consistency. Because its delivery is fairly even, Kokoro does not drift the way more expressive models do over long inputs. A ten-minute narration sounds like one take, which is harder than it sounds.
Speed. Generation is fast enough that iterating on a script is practical — rewrite a sentence, regenerate, listen, repeat. With slower engines that loop is painful enough that people stop refining.
Licensing clarity. Apache-licensed output with no usage restrictions removes the question that makes most "free" TTS unusable for monetised work.
Where it falls short
Being honest about this is more useful than a feature list.
Emotional range is limited. Kokoro reads. It does not perform. Ask it for anguish, sarcasm or building excitement and you will get competent neutral delivery with slight inflection. This is the single clearest difference against premium engines.
No voice cloning. You choose from its voice set. If you need a specific person's voice, this is not the model.
Proper nouns are hit and miss. Like every TTS model, it predicts pronunciation from spelling patterns. Unusual names, invented product names and non-English terms are unreliable. Spelling them phonetically in the input fixes most cases.
Fewer languages than commercial engines. Kokoro covers a solid core set rather than the seventy-plus that large vendors advertise. Check your language before building a workflow around it.
Running it yourself versus using it hosted
Kokoro's weights are public, so you have two real options.
Locally you download the model and run it through Python. You get unlimited generation, no network round trip and complete privacy — nothing leaves your machine. The costs are setup time, a few hundred megabytes of disk, and owning the dependency problem when something breaks.
Hosted you send text to a service and get audio back. No setup, nothing to maintain, and it works from a phone. The trade is a usage limit and trusting someone else with your text.
The honest split: if you generate speech daily, in volume, or with sensitive material, run it locally. If you generate occasionally and want to start now, use a hosted version. Most people asking about Kokoro fall into the second group and only discover they are in the first after a few months.
You can try it hosted with no account on our Kokoro TTS page — 5,000 characters a day free, MP3 or WAV download, commercial use included.
What a character allowance actually buys
Character limits are hard to reason about until you convert them to time. English narration runs at roughly 900 characters per minute, or about 150 spoken words.
| Characters | Approximate audio |
|---|---|
| 1,000 | 1 minute |
| 5,000 | 5–6 minutes |
| 10,000 | 11 minutes |
| 100,000 | Just under 2 hours |
So a 5,000-character daily allowance is around five minutes of finished audio per day — comfortable for a short video script or a set of study notes, tight for an audiobook chapter. Knowing the conversion up front saves picking a plan by guesswork.
How Kokoro compares to the other free options
Kokoro is not the only open-weight model worth knowing.
Chatterbox sits in similar territory with a slightly different character — some listeners prefer it for conversational delivery. Orpheus and Zonos are also open-weight and free here. The practical difference between them is smaller than the difference between any of them and a premium engine, so the right approach is to run your own text through each and pick by ear rather than by benchmark.
Where the real jump happens is the paid tier. If you have listened to Kokoro and specifically want more emotional range, that is the one thing more money reliably buys — the pricing page sets out what each tier costs per thousand characters.
Getting a good result out of it
Four habits improve output more than switching models.
- Punctuate for breath, not for grammar. A comma is a beat. If a sentence runs long, break it — the model paces from punctuation, so writing for the ear beats writing for the page.
- Spell difficult words phonetically. Write "Kokoro (koh-koh-roh)" once and stop fighting it.
- Write numbers as words when it matters. "Nineteen ninety-eight" removes the ambiguity that "1998" leaves.
- Generate at least a minute before judging. Twenty seconds flatters every model. Repetition in rhythm only becomes audible over a longer passage.
Frequently asked
Is Kokoro TTS free? Yes. The model is released under an Apache licence, so both the weights and the audio you generate are free to use, including commercially. Hosted services may apply their own usage limits.
How big is Kokoro? 82 million parameters, a few hundred megabytes on disk. It runs on a CPU without a dedicated GPU, which is unusual for a model of this quality.
Can I use Kokoro audio in monetised videos? Yes. The Apache licence places no restriction on commercial use of the output. Check the terms of any hosted service you use, since the service can add restrictions the model does not have.
Is Kokoro better than ElevenLabs? No, and it is not trying to be. ElevenLabs offers greater emotional range and voice cloning. Kokoro offers most of the quality at roughly a fourteenth of the running cost, with no licensing restrictions. For neutral narration the difference is small; for performance the gap is real.
Does Kokoro support voice cloning? No. You choose from the voices it ships with. Voice cloning requires a different class of model.
What languages does Kokoro support? A core set rather than the very wide coverage large commercial vendors advertise. Verify your specific language before building a workflow around it — you can check which voices exist for each language in the voice library.