Kokoro TTS: The 82M Open Model, Explained
Kokoro TTS is an 82M-parameter, Apache-licensed open model with 54 voices in 8 languages. See how it runs, what it costs to host and where it falls short.
Kokoro TTS is Kokoro-82M, an open-weight text to speech model with 82 million parameters, released under Apache-2.0. Its v1.0 release has 54 voices across 8 languages on the model card, and it runs on CPU or GPU through community tools. It is small and cheap to run, but it documents no voice cloning and no expressive tags.
What is Kokoro TTS?
Kokoro is an open-weight model published on Hugging Face. Its model card describes it as "an open-weight TTS model with 82 million parameters" and says it is "significantly faster and more cost-efficient" than larger models, with quality "comparable to larger models". Those are the author's claims. The card gives no benchmark numbers to check them against.
According to the card, the architecture is built on StyleTTS 2 and ISTFTNet. It is decoder only, with no diffusion, and no encoder was released. Output is 24,000 Hz. There are two public versions:
- v0.19, published 2024-12-25: one language, 10 voices and under 100 hours of training data. Its weights live in a separate legacy repository.
- v1.0, published 2025-01-27: 8 languages, 54 voices and a few hundred hours of training data.
The card lists 500 A100 80GB GPU hours for each version, about $1,000 of compute in total at the rates it quotes.
Kokoro specs at a glance
| Attribute | Kokoro-82M v1.0 | |---|---| | Parameters | 82 million | | License | Apache-2.0 | | Languages | 8 on the model card (the README lists 9 codes because British English is separate) | | Voices | 54 | | Training data | A few hundred hours | | Output sample rate | 24,000 Hz | | Architecture | StyleTTS 2 and ISTFTNet, decoder only |
Kokoro voices and languages
The voice list breaks down as follows. American English has the most voices, and the count of language groups is 9 because American and British English are listed separately, while the card counts English once.
| Language | Female | Male | Total | |---|---|---|---| | American English | 11 | 9 | 20 | | British English | 4 | 4 | 8 | | Japanese | 4 | 1 | 5 | | Mandarin Chinese | 4 | 4 | 8 | | Spanish | 1 | 2 | 3 | | French | 1 | 0 | 1 | | Hindi | 2 | 2 | 4 | | Italian | 1 | 1 | 2 | | Brazilian Portuguese | 1 | 2 | 3 | | Total | 29 | 25 | 54 |
Voice names begin with a two-letter code: the first letter is the language and the second is the gender, so af_heart is American English female. Quality is uneven by design. VOICES.md grades each voice. The grade for af_bella is A- and it was trained on 10 to 100 hours of audio, af_heart is graded as target quality A, and several voices are graded C or D and trained on only minutes of audio. Pick voices by grade, not by name.
How to run Kokoro locally
There are four common routes, and each one has its own install path:
- Python:
pip install kokoro>=0.9.4 soundfile, plus espeak-ng for out-of-vocabulary fallback. Create aKPipeline(lang_code='a')and choose a voice such asaf_heart. Japanese needsmisaki[ja]and Mandarin needsmisaki[zh]. - Apple Silicon: set
PYTORCH_ENABLE_MPS_FALLBACK=1to use the GPU. - Docker (community): Kokoro-FastAPI, an Apache-2.0 wrapper, serves an OpenAI-compatible API at
http://localhost:8880/v1with the model namekokoro. Images exist for CPU and NVIDIA GPU, and voices can be blended, such asaf_sky+af_bella. Output formats include mp3, wav, opus, flac, aac and pcm, and streaming is supported. - ONNX (community): kokoro-onnx, MIT licensed, describes itself as "~300MB (quantized: ~80MB)". The JavaScript package
kokoro-jsinstalls withnpm i kokoro-js. The ONNX build page lists the file sizes below.
| ONNX build | Size | |---|---| | fp32 | 326 MB | | fp16 | 163 MB | | q8 (8-bit) | 92.4 MB | | q8f16 (mixed) | 86 MB | | uint8 | 177 MB | | q4 | 305 MB | | q4f16 | 154 MB |
The ONNX page does not say which devices kokoro-js supports, so do not assume browser GPU support.
How fast is Kokoro?
The speed numbers in circulation come from community tests, not the model's authors. The Kokoro-FastAPI project reports "35x to 100x" realtime on an RTX 4060 Ti 16GB with 64 GB of RAM. In one test, a full book of about 508 minutes of audio rendered at 45.7x realtime. Streaming first-token latency was about 300 ms on GPU, about 3,500 ms on an older i7 CPU, and under one second on an M3 Pro CPU. The loaded model used roughly 3.1 to 4.0 GB of GPU memory in that setup.
Treat these as hardware-specific results. The 100x figure is an upper end on one GPU, not a property of the model. Your speed will depend on your hardware, your text and your settings.
How good is Kokoro?
On the Artificial Analysis Speech Arena, read on 2026-10-17, Kokoro-82M v1.0 scored an Elo of 1061 with a ±11 interval, from 5,374 samples. The rank range is 52 to 63. The top hosted model, Eleven v4 Turbo, scored 1330. Kokoro is good for its size and cost, but it is not near the leaders on listener preference. Another snapshot of the same leaderboard read the model between 1055 and 1066, so quote the score with its date.
What Kokoro does not do
- No documented voice cloning. The card says no encoder was released and documents no cloning method. Do not assume the model can copy a voice.
- No expressive tag controls. No emotion or paralinguistic tags are documented, so Kokoro is not marketed as an expressive model.
- Uneven voices. Several voices were trained on minutes of audio and are graded C or D.
- Eight languages on the card. Coverage is narrower than the multilingual models in the open source roundup.
What it costs to run
Self-hosting has a fixed cost. Hosted routes charge per character. The table compares the published per-million-character prices that matter most for Kokoro.
| Route | Basis | Price | |---|---|---| | DeepInfra hosted Kokoro-82M | Per 1M characters | $0.99 | | Artificial Analysis listing for Kokoro (provider price) | Per 1M characters | $0.70 | | OndaLabs Starter, Onda Lite (derived from plan prices) | Per 1M characters | $52.94 | | OndaLabs Starter, Onda Flash (derived from plan prices) | Per 1M characters | $79.41 |
The first two are third-party hosting prices, not fees set by the model's authors. The OndaLabs rows are derived: the Starter plan is $9 for 170,000 credits, and Onda Lite uses 1 credit per character. OndaLabs is a studio product, not a model you can host, so the comparison shows the scale of the difference rather than a like-for-like match. The model card's own price note dates from April 2025 and should not be quoted as current. For a worked break-even calculation, see self hosted text to speech vs hosted APIs.
Watch for look-alike sites
The model card warns that sites with "kokoro" in their root domain are not affiliated with the model or its author, and it names two examples. Download the weights and code only from the official Hugging Face and GitHub pages linked in this article.
When a hosted studio is simpler
Kokoro is the do-it-yourself end of the spectrum. OndaLabs is the no-setup end. It offers text to speech, dialogue, speech to text, custom voices, image generation and music generation in one studio. It has 30 prebuilt voices usable in every supported language, and 37 languages have dedicated pages. Delivery is controlled by describing it in plain words and with supported expressive tags, which Kokoro does not document.
The free plan is $0 with 2,000 credits per month, about 2 minutes of speech, and it includes Onda Lite, expressive tags and speech to text. Paid plans are listed on the pricing page. Languages are covered on the languages page.
Frequently asked questions
What is Kokoro TTS?
Kokoro TTS is an open-weight text to speech model called Kokoro-82M. It has 82 million parameters and is licensed under Apache-2.0. The v1.0 version has 54 voices across 8 languages on its model card.
Is Kokoro TTS free for commercial use?
The model card lists Apache-2.0, which permits commercial use under that license. Read the license file yourself before you ship. Hosted Kokoro prices from third parties are hosting fees, not license fees.
How many voices and languages does Kokoro support?
The v1.0 model has 54 voices, with 29 female and 25 male. The card lists 8 languages, while the README lists 9 language codes because British English is separate. Voice quality varies, so check the grades in VOICES.md.
Can Kokoro clone a voice?
No voice cloning is documented for Kokoro. The model card says no encoder was released, so do not plan a project around cloning. For a model that documents cloning from a short clip, see Chatterbox TTS.
Does Kokoro run on CPU?
Community tools make CPU use possible. Kokoro-FastAPI publishes a CPU Docker image, and the ONNX builds range from 86 MB to 326 MB. The authors do not state a minimum CPU, so test speed on your own machine before you commit.
How does Kokoro compare with ElevenLabs or Chatterbox?
On the Artificial Analysis snapshot read on 2026-10-17, Kokoro scored about 1061, Chatterbox scored 1025 and ElevenLabs' Eleven v4 Turbo scored 1330. Kokoro is far smaller and has permissive licensing, but it does not match the top hosted models on listener votes. Scores move, so check the leaderboard before you decide.