How does Orion hold conversations?
From the sound of your voice to Orion's spoken reply: hearing its name, knowing when you've stopped, and where the waiting time goes.
Talking to Orion feels simple: you say its name, ask something, and it answers. Underneath, it has to solve a run of small problems in order. Was that its name? When have you finished? How can it answer quickly? This lesson takes them one at a time.
Sound is numbers
A microphone measures air pressure thousands of times a second. Each measurement is one number, a Sample (audio) One measurement of the sound wave. Microphones take thousands every second; the list of them is the recording., and the list of them is the recording. Everything Orion does with your voice, it does to that list of numbers.
On Orion Orion’s microphone board does some of the work itself: it cleans up your voice and removes the sound of Orion’s own speaker.
Never missing the start
Orion only reacts after it’s sure it heard its name, and by then you’ve already said it. If it only started recording at that point, it would cut off the start of your sentence.
So it always keeps the last few seconds of sound in memory, overwriting the oldest as it goes. When the Wake word A phrase a voice assistant always listens for, like “Hey Orion”. Everything else it hears is ignored. is recognised, the moments just before it are still there to use.
On Orion Orion keeps 3 seconds in memory, and all 3 go into the recording.
Orion always keeps the last 3 seconds of sound in memory, and throws away anything older.
Hearing its name, twice
Listening for “Hey Orion” all day has to be cheap, so the first check is a small, fast detector that compares sound patterns. It catches the real thing, but also things that merely sound similar.
So anything it flags gets a second check: Speech recognition Software that turns recorded speech into written words. Orion uses a model called Qwen. writes down the first words, and Orion checks whether they actually say “Hey Orion”. Only then does it light up and listen.
Guess first Why not use only the second, more accurate check?
Right. Turning speech into text costs far more computing than matching a sound pattern. The cheap check runs all the time; the expensive one runs only when there’s a reason to.
Under the hood
The first stage is openWakeWord, which scores every 20 ms of sound against “Hey Orion”, all the time, even while
Orion is talking. The second stage asks Qwen to transcribe the 3 seconds before the detection plus the next 300 ms, and
looks for the phrase in the first eight words. While it does, the full command keeps recording, so the check costs no
time. A rejection cancels only that turn. See docs/voice-architecture.md.
- 1. Sounds like it? openWakeWord scores every 20 ms of sound against “Hey Orion”. Fast and always on, but it can be fooled. Maybe: check further
- 2. Says it? Qwen writes down the last few seconds and Orion checks the first words for the phrase. Heard:
hey orion
.Yes: Orion wakes
When have you stopped talking?
Orion has to decide when you’re done. A Voice activity detection Software that decides, moment by moment, whether a sound is someone speaking or just background noise. scores each moment from 0 to 1: how much it sounds like speech. Orion counts speech when the score goes above one line, and silence only when it drops below a lower one.
The gap between the two lines stops tiny wobbles from flipping it back and forth. After a stretch of silence, the recording ends.
On Orion Orion starts counting speech above 0.5, stops below 0.35, and ends the recording after 1.2 seconds of silence.
Cutting the silence
People pause while they think. Before the recording goes to Speech recognition Software that turns recorded speech into written words. Orion uses a model called Qwen., Orion trims long pauses down. The Model (AI) A program whose behaviour was learned from examples rather than written by hand, such as one that recognises speech. then has less audio to work through, so the reply comes sooner, and nothing you said is lost.
On Orion Pauses longer than 600 ms become 320 ms. The speech itself is untouched and stays in order.
Words, a reply, and a voice
Three programs do the main work. Qwen turns the recording into text on the lamp. Codex, an online Language model An AI model that reads text and writes text. Orion's replies are written by one called Codex., reads it and writes a reply. Piper, Speech synthesis Software that turns written words into a spoken voice. Orion uses one called Piper. on the lamp, turns the reply into Orion’s voice.
To cut the wait, Orion doesn’t wait for the whole voice. It starts speaking as soon as the first piece is ready and keeps making the rest while it talks.
On Orion If the voice is being made slower than it’s spoken, Orion switches to waiting for all of it, so it never stops mid-sentence.
Under the hood
The coordinator splits the reply text into bounded synthesis jobs. Piper returns numbered audio chunks with an end marker, which the coordinator checks before uploading them to the runtime for playback. Once a reply has switched to complete buffering, it stays that way even if later chunks come faster.
Your turn again
A conversation isn’t one question. After Orion finishes speaking, it waits for the room to go quiet, then listens for a few seconds without needing its name, so you can just reply.
Sounds and light tell you which stage it’s in, so you never have to guess whether it’s listening.
On Orion The follow-up window lasts 5 seconds and shows as a teal pulse on the light.
Under the hood
Each wake gets a fresh session ID, and a follow-up gets a new ID linked to the finished turn, so a late result from an old turn can’t sneak into a new one.
Orion listens only for its name. Nothing else you say is recorded or sent anywhere.
Interrupting Orion
Sometimes you want to cut in. But while Orion talks, its microphones hear its own voice loudest, which would drown out yours. Orion’s microphone board removes its own voice from what it hears, which is called Echo cancellation Removing a device's own speaker sound from what its microphones hear, so it can still hear you while it talks., so a “Hey Orion” said over the reply still gets through.
Because the detector listens all the time, hearing its name stops the reply straight away. Then the usual check decides: a real “Hey Orion” starts a fresh turn; a false alarm is thrown away.
On Orion ”Hey Orion, stop” just stops Orion talking. Any other request after its name starts a new conversation.
Under the hood
A detection while Orion is speaking or still working on a reply interrupts that turn, and the old reply stops before
the new command runs. Interruptions skip the usual turn toward the speaker, and an alarm being dismissed takes priority
over a new conversation. See “Wake-phrase interruption” in docs/voice-architecture.md.
- Heard over the reply?Yes: with Orion's own voice removed, the other voice stands out.
- Then: is it really “Hey Orion”?The reply stops, the check confirms the name, and a fresh conversation starts with your question.
What stays on the lamp
Your voice is turned into text on the lamp itself. What goes online is the text Codex needs to write a reply, and a photo when you ask Orion to look at something (in Demo mode, also one with each remark Orion makes).
On Orion Orion’s behaviour journal records what it did, but never camera pixels or microphone audio.
Never leaves the lamp. Speech recognition runs on the Pi and passes on only the text. A cancelled or rejected recording is thrown away.