What it is
Kokoro is an open-weights text-to-speech model with 82 million parameters, built on the StyleTTS 2 architecture by a developer known as hexgrad. Version 1.0, published on January 27, 2025, ships 54 preset voices across 8 languages, and the Python package covers American and British English, Spanish, French, Hindi, Italian, Brazilian Portuguese, Japanese and Mandarin.
The weights are Apache 2.0. The model card says it was trained on a few hundred hours of public-domain and permissively licensed audio plus synthetic speech generated by closed TTS models from large providers, for about $1,000 of GPU time. Its Hugging Face page records more than 10 million downloads in the past month. There has been no new model since v1.0, and the GitHub repository's last commit was in August 2025.
Our review
Kokoro is the default answer when speech has to be cheap, local and legally simple, and the wrong one when it has to sound human. Artificial Analysis ranks it 57th of 98 text-to-speech models at 1065 Elo, level with Amazon Polly Generative (1066) and ahead of the open Chatterbox (1024), at $0.65 per million characters, the lowest price on the board. Eleven v4 Turbo leads at 1335 and costs $40 for the same million characters.
Size explains the economics. The whole repository, weights and voice packs together, is about 330 MB, and the model is a decoder-only StyleTTS 2 design with an ISTFTNet vocoder and no diffusion step. The model card's rule of thumb is 1,000 characters to a minute of audio, which puts hosted Kokoro at under $0.06 an hour of speech.
The licensing is simpler than most, though not quite as clean as "permissively licensed audio" implies. The weights are Apache 2.0. The training set mixed public-domain, Apache- and MIT-licensed and CC BY audio with synthetic speech from closed commercial TTS models, which the card defends with US Copyright Office guidance. Whether that sits comfortably with those providers' terms is a question for counsel rather than the model card.
Development has stalled. v1.0 shipped on 27 January 2025, the repository was last touched in August 2025, and there is still no cloning or style control beyond 54 preset voices. Its popularity has drawn imitators too: the card warns that websites with "kokoro" in the domain are not affiliated and are likely scams, a real risk for buyers looking for a hosted version.
Where it earns its place
An accessibility engineer at a regional news publisher
Add listen-to-article audio to every story without a large per-character bill
Pre-render each article on existing servers, or call a host: at $0.65 per million characters, 50 million characters a month is about $33. Listeners get a clear but plainly synthetic voice.
A developer of an offline reading app for Android and desktop
Ship built-in narration that works without a network
The small download and Apache licence allow bundling without royalties. The package relies on espeak-ng for some languages and unfamiliar English words, so that dependency and its licence travel with the app.
A learning-and-development manager with 2,000 internal training modules
Voice low-stakes course content cheaply
100 million characters, roughly 1,700 hours of audio, is about $65 on hosted Kokoro against about $4,000 on Eleven v4 Turbo at Artificial Analysis prices. Learners will hear the 270-point Elo gap.
Before you commit
- Setup
- pip install kokoro (0.9.2 or later) plus espeak-ng, then a few lines of Python with KPipeline; output is 24 kHz audio. Apple Silicon GPUs need PYTORCH_ENABLE_MPS_FALLBACK=1.
- Integrations
- Python package built on the misaki grapheme-to-phoneme library; hosted endpoints on Replicate and DeepInfra; a demo Space on Hugging Face.
- Data and privacy
- Apache 2.0 weights, commercial use allowed. Training mixed public-domain, permissively licensed and CC BY audio (Koniwa, SIWIS) with synthetic output from closed TTS models. Self-hosted, text stays on your servers.
- Cost at scale
- $0.65 per million characters on Replicate (by Artificial Analysis's figure) and $0.99 on DeepInfra, so 100 million characters a month, about 1,700 hours of audio, costs $65 to $99 hosted.
- Limits
- 54 preset voices in 8 languages, no voice cloning or emotion control, 1065 Elo, and no model release since v1.0 in January 2025.
What's good
- Apache 2.0; about 330 MB including voice packs
- $0.65 per million characters hosted, cheapest on Artificial Analysis
- 54 preset voices across 8 languages
- Decoder-only design with no diffusion step
What's not
- 1065 Elo: level with Polly Generative, 270 below the leader
- Preset voices only; no cloning or emotion control
- No model update since January 2025
- Training data includes synthetic audio from closed TTS models
Best for
Developers who need cheap, permissively licensed speech for narration, accessibility or prototypes.
Alternatives
Kokoro-82M: common questions
- Is Kokoro-82M worth it?
- Nothing this small sounds this good, but mid-table quality, preset voices and a model unchanged since January 2025 cap its ambitions. We rate it 3.9 out of 5.
- How much does Kokoro-82M cost?
- Free under Apache 2.0. Hosted at $0.65 per million characters on Replicate (via Artificial Analysis), the cheapest on its leaderboard, and $0.99 on DeepInfra.
- What are the downsides of Kokoro-82M?
- 1065 Elo: level with Polly Generative, 270 below the leader; preset voices only; no cloning or emotion control; no model update since January 2025; training data includes synthetic audio from closed TTS models.
- Who is Kokoro-82M best for?
- Developers who need cheap, permissively licensed speech for narration, accessibility or prototypes.
- How much does Kokoro-82M cost for a team or in production?
- $0.65 per million characters on Replicate (by Artificial Analysis's figure) and $0.99 on DeepInfra, so 100 million characters a month, about 1,700 hours of audio, costs $65 to $99 hosted.
- Is Kokoro-82M safe for company data?
- Apache 2.0 weights, commercial use allowed. Training mixed public-domain, permissively licensed and CC BY audio (Koniwa, SIWIS) with synthetic output from closed TTS models. Self-hosted, text stays on your servers.
- What are the best Kokoro-82M alternatives?
- The closest alternatives we have reviewed are Chatterbox, Qwen3-TTS and Gemini 3.8 Flash TTS.
Answers checked Oct 23, 2026.