Yumii's voice is where "local-first" is most visible: the default engine runs entirely on your CPU and needs no key. Pick the provider in Dashboard → Settings → Voice (TTS).
Kokoro — the local default
Kokoro-82M is a small, remarkably natural neural voice that Yumii runs via ONNX on the CPU:
- Downloads once on first launch (~330 MB, with a progress bar in the
orb) to
~/.yumii/models/— after that it works fully offline, forever. - 54 voices — the default is
af_heart; the dashboard offers a curated picker, and any voice name can be set viaKOKORO_VOICEin the config. - Fast enough to feel live. Yumii splits her reply at sentence and clause boundaries and synthesizes incrementally, so the first words reach your ears in about a second while the rest is still being generated.
- English-first. Other languages exist in the voice bank but the quality varies; for serious multilingual use, a cloud voice is better today.
fp32 vs int8
Kokoro has a smaller int8 variant, and counterintuitively it's about 3–4×
slower than the full fp32 model on ordinary x86 CPUs (quantized ops
fall back to slow kernels). Yumii defaults to fp32; leave
KOKORO_MODEL_SIZE alone unless you've measured otherwise on your
hardware.
ElevenLabs — the cloud premium
The most expressive and lifelike voices available, streamed from the cloud.
Needs an ElevenLabs API key and a Voice ID (pick any voice at
elevenlabs.io/voice-library and copy its ID — it looks like
21m00Tcm4TlvDq8ikWAM). Costs per character after the free allowance; every
word she speaks is a network call.
CAMB.ai — the cloud alternative
A high-quality streaming alternative with strong multilingual coverage. Needs a CAMB API key and a numeric Voice ID from client.camb.ai.
Choosing
| Kokoro (default) | ElevenLabs | CAMB.ai | |
|---|---|---|---|
| Runs | Locally, offline | Cloud | Cloud |
| Cost | Free forever | Paid after free tier | Paid after free tier |
| Latency | ~1s to first word | Network + synth | Network + synth |
| Expressiveness | Very good | Best-in-class | Very good |
| Languages | English-first | Broad | Broad |
| Privacy | Nothing leaves your machine | Text of her replies goes to the provider | Same |
The honest recommendation: stay on Kokoro unless you have a specific reason not to. It's the only option that keeps her voice private, free, and independent of your network — and for a companion you talk to daily, those three compound.
Provider and voice changes need a restart (quit from the tray, reopen); the dashboard reminds you after saving.