rhapsode-engine-kokoro

Kokoro as a rhapsode engine, through kokoro-onnx. It runs on the CPU and brings no torch, which makes it the engine to install first: no GPU, a download measured in megabytes, and speech faster than realtime on a laptop.

pnpm wizard install kokoro

It needs Python 3.11 to 3.13: 3.11 for the SDK, and below 3.14 because kokoro-onnx says so. With uv on the path the installer fetches one; without it, set install.python to an interpreter in that range.

What it can do

Three variants, which are the same model at three precisions: fp16 (the default), fp32 and int8. They claim the same things: no cues, no deliveries, one dial, speed (0.5 to 2.0), which is also what the OpenAI shim's speed maps to. English only for now, in 28 voices; each voice's first letter is its accent, a American and b British.

It does not clone from audio, but it makes voices two other ways (docs/protocol.md § 7):

• A blend: blend=af_bella(2)+af_sky(1), Kokoro-FastAPI's recipe, mixed once and stored under the id you give it. It reads the 28 MB voices file and never the model.
• A style vector as the reference: a .npy of shape 510 x 1 x 256, or a Kokoro-FastAPI .pt voicepack, which is how its v0 voices (am_v0gurney and the rest) come across. A .pt is read as data and never unpickled. A pack named the Kokoro way (bm_...) keeps its British accent.

Licences

The weights are Apache-2.0 and commercially usable. The code the worker runs is GPL-3.0-or-later, because kokoro-onnx turns text into phonemes with phonemizer and eSpeak NG. This package is MIT, and the catalog reports the GPL, because that is the licence a commercial user needs to know about.

What it costs

Measured on an Apple Silicon Mac, on the CPU, September 2026, for a 6.2 s line:

Variant	Download	Load	Speed
fp16	177 MB	0.31 s	9.9x realtime
fp32	326 MB	0.48 s	8.3x realtime
int8	92 MB	0.25 s	2.5x realtime

Through the core, fp16 speaks at 8.5x to 9.2x realtime once warm, and a first /speak (spawn, load, speak) took 1.75 s for 6.2 s of audio. The adapter passes all 27 conformance checks against the real weights, on all three variants. That run was on Python 3.14 with the version cap overridden by hand, because no 3.13 was to hand; the installer will not do that for you.

Every variant also needs the 28 MB voices file. The weights land in ~/.cache/rhapsode/kokoro, or RHAPSODE_KOKORO_WEIGHTS, and are checked against a SHA-256 as they arrive.

One thing to know

eSpeak NG reads its data from a path inside the engine's virtualenv, and exits the whole process if that path is 160 characters or longer. The adapter refuses to load rather than let that happen, and says so; the cure is a shorter install.venvDir.