Audio Gen · Model Guide · 4 min read

Kokoro TTS

An 82-million-parameter TTS that runs on a CPU and sounds shockingly good — local speech with zero friction.

Basic Information

Kokoro is the little miracle of local audio: at just 82 million parameters — a rounding error compared to the billion-parameter models elsewhere on this page — it produces natural, expressive English (and multiple other languages) speech that repeatedly beats far larger models in community blind comparisons. It is fully open under Apache 2.0, runs in real time on an ordinary CPU, and needs under a gigabyte of memory.

For local-first projects it is the default answer for speech: private, free, fast, and good enough that many listeners cannot tell it is synthetic.

Best use cases

• Fully offline TTS for apps, kiosks, and embedded devices — no API, no fees
• Audiobook and article-reading pipelines on any hardware
• Voice interfaces for edge devices and air-gapped systems
• Batch narration where cloud per-character pricing would sting

Computational Requirements

Trivial: under 1 GB, and it does not require a GPU at all — it runs real-time on desktop CPUs, which makes it the most accessible audio model in this list.

Commercial license

Apache 2.0 — unrestricted commercial use, embedding in products, and redistribution. Voice packs carry their own attribution notes; check the model card for the voice pack you pick.