Voice¶
Engineering → Starbase → Voice: how the AI hears and speaks. Speech is turned into text on the Starbase computer, in its own process: if it ever crashes, chat keeps working.
Voice files¶
What hearing, the AI's voice, voice ID and wake words run on. Downloaded once from each project's own release,
checked before use, and kept on the Starbase computer (data\models).
| Pack | What it's for | Size |
|---|---|---|
| Hearing: Parakeet | speech to text, the most accurate | 480 MB |
| Hearing: Moonshine | speech to text, lighter and faster (for slower PCs) | 250 MB |
| The AI's voice: Kokoro | how the AI speaks | 354 MB |
| Voice ID | knows who's talking | 27 MB |
| Wake words | hears "Hey Celestra", "Hey Ship"… | 35 MB |
| Speech detector | tells people talking from noise | 5 MB |
One hearing model is enough. Download what's missing gets the rest.
Wake words¶
Every AI's name with "Hey" or "OK", "Hey Ship", and a distinctive name on its own when the persona allows it. Sensitivity: lower it if devices wake by themselves, raise it if they miss you. Devices pick up changes at once.
Listening¶
What starts and ends what's heard, for every device (the Bridge and the room devices):
- Wait for them to start: how long after the wake word it waits for you to start talking (8 seconds to start with, so you can stop and think).
- Pause that means they're done: how long a pause ends what you're saying.
- Keep listening after it answers: how long you can keep talking afterwards without its name.
- Listen for: Speech (people talking start and end what's sent, not background noise) or Any loud sound.
- Speech detector: Low (only clear speech), Normal, or High (hears quieter speech).
Voice ID¶
Each person teaches the AI their voice once (Ready Room → Your AIs → Voice: three sentences). Voice match is how close a voice must be to count as them; the list shows each person's voiceprint, and the Captain can delete one. Voiceprints never leave the Starbase computer.
Devices are paired in the Shuttlebay.