Choose an AngeVoice model

Select a model for the task first. No model is the best option for every host and workflow.

KokoroZipVoiceMOSS

Model comparison

Model	Best for	Advantages	Considerations
Kokoro	Daily Chinese speech, agent responses and low-latency synthesis	Lightweight, preset voices and the default recommendation	Does not clone from reference audio
ZipVoice	Custom voices, long text and high-quality cloning	Saved Voice Profiles and reference text support	Higher cold-start and inference cost; reference audio should be short and clean
MOSS-TTS-Nano	Reference-audio cloning and varied speech tasks	Temporary or server-side reference audio	More sensitive to reference quality and runtime tuning

Choose by goal

Goal	Start with
Reader, notification or conversational agent speech	Kokoro
Clone a short reference voice	ZipVoice or MOSS
Long narration with a custom voice	Test ZipVoice, then compare MOSS
Low-power NAS	Kokoro, with cloning models loaded only when required
GPU with limited VRAM	Load one active model and allow CPU fallback for others

Reference audio and Voice Profiles

Use one clean speaker without music, reverb or clipping.
ZipVoice works best with a short reference. Longer input is accepted but may reduce speed or quality.
A saved Voice Profile stores the reference WAV and matching transcript.
Browser recording requires HTTPS or localhost.

Verify a model switch

curl http://127.0.0.1:8100/v1/models
curl http://127.0.0.1:8100/v1/models/current

Check the actual provider and fallback reason. The first load may be slow; warm requests are more representative of everyday latency.

Memory and VRAM

The default configuration keeps only the required model loaded and can unload it after idle time. Preloading all three models is rarely useful for a personal deployment.