Kokoro-82M is an 82-million-parameter, Apache-2.0 text-to-speech model with 54 voices across 8 languages, small enough to run on modest hardware and cheap to host. Pick it when you want self-hosted speech output with no per-character API bill. It's not built for expressive voice cloning; check pronunciation and audience fit before using it for anything public.
Other good fits
A project that needs a permissively licensed (Apache-2.0) voice model with no per-character API cost.
+ Pros
Read pros
Kokoro-82M is a small, open-weight text-to-speech model (82 million parameters, Apache-2.0 license) with 54 voices across 8 languages as of its 1.0 release. Because it's small, it runs on far more modest hardware than larger voice models and processes audio at 24kHz. It was trained only on permissively licensed or machine-generated audio, so its own output doesn't carry someone else's copyrighted voice.
Pick it when…
You want to self-host text-to-speech and avoid ongoing per-character API charges.
Your hardware budget is modest; Kokoro's small size means it doesn't need a large GPU.
You need a clearly licensed (Apache-2.0) model for a project with its own licensing requirements.
− Cons
Read cons
Kokoro's documentation doesn't publish a guaranteed speed or memory minimum for every device, so test on your actual target hardware rather than assuming it will be fast enough. It's built for efficient, natural narration rather than highly expressive or emotionally varied speech. Every generated clip still needs a human check for pronunciation, accessibility and whether it fits the intended audience before it goes out publicly.
Pick something else when…
You need expressive, emotionally varied speech.F5-TTS is built around a more expressive synthesis workflow; check its license before use.
You need voice cloning from a short sample.XTTS-v2 is designed for that, with its own consent and licensing considerations.
You'd rather not host anything yourself.OpenAI TTS gives you a managed API with no server to run, at a per-character price.
Avoid if
You'd rather not manage hosting; a managed API is simpler for occasional use.
Context
The current release is version 1.0 (January 2025), with 54 named voices across 8 languages. Pin the exact release and voice you're using in your own documentation, since community wrappers and similarly named projects can differ from the official hexgrad/Kokoro-82M weights.
Requirements
Install via pip locally, or use a hosted inference provider (Replicate, DeepInfra are documented options) if you'd rather not run it yourself. It needs espeak-ng for phoneme processing and the misaki library for text handling. No specific memory or latency figures were measured for this review; test on your own hardware before committing.
Licence & cost
The model weights are free under Apache-2.0. If you self-host, you pay only for the machine: Kokoro's small size (82M parameters) means it runs on modest hardware, well below what larger voice models need. If you use a hosted inference provider instead, expect API pricing around or under $1 per million input characters, though exact current rates should be checked with that provider. Ongoing effort is minor: no fine-tuning is required for basic use, and updates are just a package upgrade.
First task
Documentation-based suggestion — not run here.
Record a short accessibility notice.
Install Kokoro locally, or use one of the documented hosted providers.
Write a short, factual accessibility notice (2-3 sentences) with any tricky names or numbers spelled out phonetically if needed.
Generate audio with one of the 54 voices in your target language.
Have a colleague listen and compare the audio against the written text, checking pronunciation.
Done when a listener who hasn't seen the text can write down what was said correctly, including any names or numbers.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on the Kokoro-82M model card and GitHub repository, checked 23 September 2026. No installation, generation or listening test was performed for this review.
Kokoro-82M model card · Sep 23, 2026 — 82M parameters, Apache-2.0 license, 54 voices/8 languages, training data scope, hosted provider pricing note.
Kokoro repository · Sep 23, 2026 — Installation (pip), dependency on espeak-ng and misaki.
Kokoro voices · Sep 23, 2026 — List of the 54 named voices and their languages.
Further review notes & alternatives
Alternatives and trade-offs
XTTS-v2 wins when you need voice cloning from a sample rather than a fixed voice list.
F5-TTS wins when expressive, varied delivery matters more than Kokoro's small footprint.
OpenAI TTS wins when you'd rather pay per character than manage hosting yourself.
Generate speech locally with a lightweight Apache-2.0 model.
Run Kokoro on a CPU or modest GPU for embedded applications.
Choose Piper separately when Raspberry Pi efficiency is the priority.
