The details

Talking to Directive 47 instead of typing at it. Hold a key, speak, let go, and what you said is handled exactly as if you had typed it — or put your hands back on the stick and let it decide for itself when you are talking to it.

No audio and no transcript leaves your machine. Speech is turned into words by a model running on your own computer. The model file itself is downloaded once, from huggingface.co; after that, nothing about your speech goes anywhere.

Ask for it

“can you hear me” “what microphone are you using” “is my push to talk key bound twice”

Microphone: Yeti Nano, capturing.
Gate: hands free, opening when you say my name.
Right now: the microphone is open and I am waiting to hear you start.
Push-to-talk: CapsLock (hold).
Warning: CapsLock is also bound in Elite (KeyboardMouseOnly) to HeadLookToggle. One of the two
will not work, and neither will say so — pick another key for one of them.
I answer to: D47.
Echo cancellation: running, so you can talk over me.
Transcription: base.en loaded.

Everything comes back together on purpose. “It cannot hear me” has five possible causes — no key bound, no microphone, the microphone gone, the key clashing with Elite, or no speech model — and you should not have to guess which.

Getting set up

Three things, in this order:

1. Bind a key. Already done — right shift, out of the box. Clear the row if you would rather Directive 47 never opened the microphone at all.

2. Download a speech model. Also already done, or under way. Directive 47 ships with Tiny (English only) selected and fetches it from huggingface.co the first time it starts — about 75 MB, once. Until it lands, Directive 47 captures your voice and tells you plainly that it cannot understand it yet.

3. Check the key is yours alone. If Elite is already using it, say so is exactly what Directive 47 does — see below.

Optional: put the key away entirely. How D47 decides you are talking to it has two hands-free settings. They are off out of the box, deliberately — that section says what turning one on actually means.

Your key might already be Elite’s

A key bound in both places has no symptom other than not working, in one direction or the other, depending on which application sees it first. Nothing tells you; it simply does nothing.

So Directive 47 reads your actual Elite bindings — the preset you are really using, including the built-in ones — and says outright when your push-to-talk key is already spoken for:

Warning: CapsLock is also bound in Elite (KeyboardMouseOnly) to HeadLookToggle. One of the two
will not work, and neither will say so — pick another key for one of them.

And the same for a stick button, with one difference that matters. Elite records a joystick binding against its own name for the device, which is not the one Directive 47 reads — so when a button of that number is bound in Elite, it cannot tell you whether that is the same stick or a different one on your desk. It says so in those words rather than pretending to be sure:

Warning: button 7 on the Virpil Alpha may collide. Elite (Custom) binds a button of that number
to SelectTarget, and I cannot tell whether that is the same controller. If the microphone will
not open, this is the first thing to check.

Finding nothing is the stronger answer of the two, and is said plainly: no button of that number is bound anywhere in your preset, on any device, so there is nothing left to be unsure about.

Your bindings are never written to. Directive 47 only ever reads them.

If it has not managed to read them, it says nothing rather than giving you an all-clear — never having looked is not the same as having looked and found nothing.

Settings

Microphone

Which input to listen on. Leave it unset and Directive 47 follows your Default Device — the microphone Windows Sound settings shows first — not the separate Communications default some headsets split off. Left on “system default”, it keeps following that row while it runs, so switching to a headset mid-session moves the open microphone with it. A device you have chosen by name stays chosen instead.

If the one you chose disappears, the status answer names it rather than reporting generic silence. No microphone at all is a feature being off, not a failure — Directive 47 stays fully usable typed. A move that lands mid-sentence says it did not catch that rather than dropping the words silently.

Push-to-talk

What you hold to talk. Right shift out of the box — a Commander on a stick and throttle has a spare thumb and not much else, and it is the right-hand shift specifically, so the left one you may already be using in the game is not this.

One row, and it takes a key, a stick button, or both. Press Press to bind and Directive 47 listens for either at once: press a key and it takes the key, press a button on your stick and it works out which one that was.

To have both, bind twice — one gesture each time. Press the control and give it a key; press it again and give it a button. The two are stored separately, so the second does not replace the first, and the row then reads RightShift, button 11. Either one opens the microphone, and the last one you let go of closes it — so letting go of the key while your thumb is still on the button does not cut you off mid-sentence.

It is one of each, not any number. There is one slot for a key and one for a stick button. Give it a second key and the key changes; give it a second button and the button changes. Nothing is ever silently added to a list you cannot see — what the row reads is exactly what is bound, and two gestures is the most it will ever say.

A key on its own, like right shift, binds when you let it go. Modifiers are the one kind of key that cannot be taken the instant they go down, because that is also how a combination starts. Hold one and press something else and you get the combination; press one and release it and you get the modifier. Right shift is the default here for exactly that reason — it is a key nothing else wants.

Unbind clears both. That is what the word says, and nobody ends up with two by accident.

There used to be two rows for this. The stick half is described below and is unchanged in every respect except that it is no longer a separate question.

Clear the row and Directive 47 never opens the microphone — unless you have also put the row below into one of its hands-free settings, which is the reason for those and the only case where an unbound key still leaves a live microphone.

Clearing it was the old default, and it meant a voice companion that could not hear anything until you found this row.

In the two key-driven settings, nothing is kept unless the key is held.

Pressing it does not silence Directive 47, and that is deliberate. It did for one release, and the cost was that every tap meant as be quiet also recorded a second of your quiet room — which a speech model does not transcribe as nothing. Interrupting has its own control now: Cancel, the row directly below this one, which stops the voice and abandons the turn.

Unlike the stop key, this one does not need a modifier — a bare key is the normal arrangement for push-to-talk, which is exactly why the collision check above matters. If right shift is bound to something in Elite on your setup, the status report above says so by name.

The model cannot change this. A model that could unbind your microphone key has taken away how you talk to it.

The stick half

Not a row of its own any more — it is the Push-to-talk row above. This section is about what happens when you press a stick button at it, which is worth reading before you try.

Press Press to bind, then press and release the button you want, and Directive 47 works out which one it was. It is stored separately from your key, so binding a button does not unbind the key: you said two things rather than changed your mind about one.

It has to be a button that springs back. A switch that stays where you put it would hold the microphone open until you moved it again, so the capture declines one and says why. Those belong on the switch panel instead, which is the other half of the same hardware — Phase 21 turned away every springing button because a switch needs a position to mean anything, and this feature is that decision read the other way round.

If the controller is not there when D47 starts, it says so and your key carries on working. A stick that is asleep is one of the ways “D47 cannot hear me” happens with no reason attached.

The collision check is weaker here than for a key, and it says so. Elite records a joystick binding against its own internal name for the device, which is not the name Windows gives it, so Directive 47 cannot tell whether Elite’s button 24 is on the same stick as yours. What it can say is that a button of that number is spoken for somewhere — worth saying on a HOTAS, where every button is usually already used:

Push-to-talk button 24 may collide: Elite (Custom) binds a button of that number to
UseBoostJuice. D47 cannot tell whether that is the same controller.

The model cannot change this one either, for the same reason.

Cancel

Ctrl+Alt+X out of the box. One press does three things:

• stops Directive 47 talking, instantly, mid-word;
• abandons the turn it is working on — the model stops, and so does the spending. This is the one worth knowing about: a web search that is going to take twenty seconds and cost you money is stopped by this and is not stopped by saying “be quiet”;
• does both from anywhere, including while Elite has the foreground.

Bind a key, a stick button, or one of each. The button half is what makes this usable in the middle of a fight, and it is the only other binding in Directive 47 that takes one — the interface hotkeys stay keyboard-only, because you press those sitting at the window. Giving it the same kind twice replaces that one; Unbind clears both.

Claimed system-wide rather than only while Directive 47 has focus, because the moment you want it is the moment Elite is in front — so the key half needs a modifier, and a bare letter is refused as you press it. The stick half has no such rule.

Push-to-talk does not do this, and for one release it did. Pressing push-to-talk used to silence Directive 47 as well as open the microphone, which sounds tidier than it was: every tap meant as be quiet also recorded a second of your quiet room, and a speech model handed a second of silence does not return nothing — it invents a sentence, and Directive 47 answered it. Push-to-talk is push-to-talk again, and interrupting is this.

The model cannot change this row. It can be changed from the panel or by voice — but not by anything the model calls. A model able to unbind the one control that outranks it has removed the one control that outranks it, and that matters more now the same press ends the turn.

How D47 decides you are talking to it

Four settings, and the key works in all four — a policy that decides for itself is not a reason to take away the one that does not.

What the last two actually mean. The microphone was already open — it always is, so the pre-roll has something in it — but until now nothing was ever kept unless you were holding the key. In the two hands-free settings, every stretch of speech in the room is captured and transcribed before Directive 47 can decide whether it was meant for it. That is a real change and it is why these are off out of the box.

What does not change: none of it leaves your machine and none of it is written to disk. The speech model runs locally, a stretch that was not addressed to Directive 47 is discarded without reaching the transcript, and the panel says the microphone is open the whole time it is — see Seeing that the microphone is open.

The cost is CPU. Everything said near you is transcribed and then mostly thrown away, so run these on one of the smaller models unless you have cycles to spare.

The model cannot change this row. A model that could put Directive 47 into continuous listening could start capturing on your machine — and anything the model can call, a hostile in-game message can try to invoke. You can still say “listen for your name” or “stop listening all the time”, which go through the keyword router rather than through the model.

How much louder than the room speech has to be

Only applies hands free. Directive 47 measures your room continuously — falling to a new quiet almost at once, rising to a new loud slowly — so this is a margin above whatever your room happens to be, not a fixed loudness. That is what makes one number work across a headset boom, a desk condenser and a laptop array, whose levels differ by tens of decibels.

Lower hears more, and will open on a cough or a keyboard. Higher waits until you are clearly talking. 9 dB out of the box.

Quiet that ends a sentence

How long you have to stop talking before Directive 47 decides you have finished. 700 ms out of the box, which is deliberately generous: people pause mid-sentence to look at something, and an utterance cut at the first gap reaches the model as half a question. Being wrong the other way costs a little dead air on the end of the clip, which the speech model ignores.

What D47 answers to

Only applies in Listen when I say its name. Leave it unset and Directive 47 answers to whatever you call your ship’s AI, so renaming the core renames the wake word with it.

Set it — comma-separated — when the speech model keeps hearing the name as something else. It is already forgiving: the comparison ignores punctuation and spacing, so D47, D 47, d-47 and D47. are all the same word. Adding a spelling is for when it comes out as something genuinely different.

The name has to be near the front of what you said. Talking about Directive 47 is not talking to it.

Seconds D47 keeps listening after you say its name

Say the name on its own, Directive 47 sounds its listening cue, and the next thing you say is the request — the way you would address a person. 12 seconds out of the box, and the follow-up closes the window again: it is one reply, not an open microphone. Nothing is said back and nothing reaches the transcript; being called by name is not conversation.

Set it to zero if you would rather the name and the request always arrived in the same breath.

Cancel D47’s own voice out of the microphone

On, in every setting rather than only the hands-free ones — holding the key while a callout is being read out otherwise transcribes the callout.

On speakers, this is what lets you talk over Directive 47. Measured against a simulated room, it removes upwards of 25 dB of its own voice while leaving yours essentially intact.

If it cannot start — the native library missing, usually — it says so and the microphone keeps working. Hands-free listening then goes deaf while Directive 47 is speaking, rather than risk the loop where it hears itself, transcribes itself and answers itself. Ask “can you hear me” and it will tell you which of the two you are in.

On headphones none of this matters much, because there is no echo to cancel.

Take the room out of what D47 hears

On. Suppresses steady background noise — fans, a headset’s own hiss — before the speech model sees it. It also makes the hands-free decision easier, since that decision is entirely about how far a sound sits above the room.

Capture before the key

How much audio from just before the gate opened is kept. 500 ms out of the box.

It exists because the key is sampled ten times a second, so a key-down is noticed up to 100 ms after it happened — and without this the first syllable of every sentence is clipped, which is where the proper nouns are. It does the same job hands free, where what it covers is the moment before Directive 47 was willing to call the sound speech.

Names it has learned to hear

Proper nouns are where speech recognition fails hardest and least visibly. A misheard system name does not come back as an error or as a low-confidence marker — it comes back as a plausible English word, and the answer is confidently about the wrong place. Directive 47 already biases the transcriber towards the names around you, but that list holds sixty and the galaxy holds four hundred billion, so a name you have not been near lately cannot be on it.

So there is a recovery. Ask about a system it cannot find and it will not shrug politely: it looks through the places you have actually been — every system, station and faction in your own journals — for one that sounds like what it heard, offers those back, and runs your question again when you say which. Nothing is sent anywhere to do it; those names came off your own disk.

What it remembers is the word, not the answer. Correct it once about Eurebia and it has learned that word, so “the Eurybia Blue Mafia” comes out right too, and so does every other sentence containing it. An alias against the system would have fixed one question.

It only ever learns when you correct it, and it refuses to learn a word that already means something — a place you have visited, or a phrase Directive 47 answers to. A wrong entry here would be permanent, invisible and would silently rewrite everything you said afterwards, which is a worse failure than the mishearing it came from. It is also spoken input only: if you type a name, that is what you meant.

This row lists what it has learned, and Forget them all clears the lot. Everything here stays on this machine and belongs to the Commander flying — nothing is learned from a chat message, a mission description, or anything another player wrote.

Speech model

Which Whisper model turns your speech into words. The row marks which are already on disk and what the others would cost to fetch, so you can see which choices are already paid for.

Base (English only) is what a fresh install has selected. Tiny is cheaper and is offered, but it mis-hears the words Directive 47 acts on — over a recorded corpus it heard “Cancel that” as “Cancer that”, and “cancel that” is one of the phrases that interrupts it, so the model saved you 67 MB and cost you the way to shut it up.

Every model here is English-only. The multilingual ones were withdrawn: Directive 47 asks Whisper for English on every clip, so they cost the same download and gave back a model worse at the one language they were being asked for. If your settings still name one, you get its English twin automatically and nothing is lost.

What a bigger model costs is time per sentence, not time per second of audio. Whisper reads your speech in 30-second windows and pays for the whole window whether you filled it or not, so a two-second question and a twenty-second one cost the same — and the model you picked is what sets that figure:

Measured on a 24-core desktop with an RTX 5080; a smaller machine is proportionally slower, and anything over 30 seconds in one breath costs another window. That is the whole delay between you letting go of the key and Directive 47 starting to think — so if it feels slow to answer, this row is the first place to look.

Medium is the most accurate and effectively wants a GPU. It was the only model that heard “Deciat” every time it was said, where the others offered “DCI” and “DC at” — proper nouns are where the difference shows, and they are most of what you say to a ship’s computer. Three seconds a sentence on the CPU is the price without one.

The device never changes the words. Running on the GPU is a speed and memory choice only: across a 37-clip corpus, every English model produced byte-identical transcripts on CPU and GPU.

Choosing a model downloads it. Selecting one you do not have starts the transfer there and then, with the size and progress on the row; the same thing happens at startup for a model that is selected and missing. The choice is the go-ahead — the size is on the row before you make it, and huggingface.co is listed under Privacy for as long as a model is selected.

Directive 47 knows what each model file should be. The SHA-256 of every model it offers is written into the build, and a download that does not match it is discarded rather than loaded. That is worth one sentence about what it actually buys: those values were read from huggingface.co once, on a stated date, and pinning them does not make that first read trustworthy — it means the file changing afterwards becomes visible, where before the expected hash and the bytes came from the same place. The model is loaded and run on your machine, so it is worth checking.

none stays a real choice. Pick it and Directive 47 hears you and says plainly that it cannot turn what it heard into words.

Running on the GPU

Off by default, and about five times faster when you turn it on. Measured on an RTX 5080 with Small (English only): 190 ms against 920 ms for the same sentence.

It costs video memory, which is the part worth thinking about, because that memory comes out of whatever else wants the card. It is close to the size of the model file, because the weights are almost all of it:

In VR that trade is a real one. Your GPU is already the scarce thing there, and taking memory and time from the game shows up as dropped frames rather than as anything that looks like a speech problem — a symptom nowhere near its cause. That is why this is off out of the box rather than on. On the desktop window, with headroom to spare, it is close to free.

If your machine has no GPU D47 can use, it runs on the CPU and says so — in the log, and in what it reports about itself. It does not fail, and it does not claim a device it is not using.

The switch takes effect immediately, both ways: turn it off and the video memory is handed back.

This did not work before #187. The toggle shipped for months with no GPU code behind it at all — the CPU runtime accepted the request, loaded without error, and the log read “on the GPU” because it was repeating the request back rather than reporting what happened. If you turned it on and noticed nothing, that is why. D47 now uses Vulkan, which works on AMD and Intel cards as well as NVIDIA.

Seeing that the microphone is open

Bottom left of the panel, on the desktop and in the headset, in mini as well as full. Three states:

The first two name the mode because the state only happens in that mode: at rest is push-to-talk, hands free is not. The open gate is reached both ways — a held key and a gate Directive 47 opened for itself are the same fact about the microphone — so it says neither.

Filled or hollow is the state, not only the colour — a glance reads a shape first, and a difference that is only colour is not a difference for everybody. Hover it for the gesture, or the name, that would open the gate.

Nothing is drawn when no device is open at all.

Downloading a model

A selected model that is not on disk is fetched — at startup, or the moment you choose it. There is no prompt to answer: the selection is the go-ahead, the size is on the row before you make the choice, and the shipped default is the smallest model in the list.

The download is checked against the checksum huggingface.co publishes for the file, as it arrives, and thrown away rather than kept if it does not match: a file that fails its checksum is either a broken transfer or something that should not be loaded, and both answers are the same. The size Directive 47 reports is the one the host actually gave, not a figure written into the app that would go stale the first time a model was republished.

If the download fails — no network, the host refusing — the selection stays where it is and Directive 47 says it has no speech model loaded when you ask. It tries again the next time it starts. Choose none if you would rather it stopped trying.

Spelling a value onto a keyboard

Wherever Directive 47 draws a keyboard — the panel’s own, and the one a controller ray opens on a text box in the headset — you can say the value instead of pointing at every key.

Say each letter as its word:

“alpha bravo seven done”

That types ab7 and presses Done, in one breath. Done is a key, so an utterance can spell a value and commit it together. So are delete, clear and cancel.

The words are the NATO alphabet, which is the one spelling alphabet a Commander in Elite already knows. It is also the one Whisper transcribes reliably: single letters come back as “bee”, “sea” and “you”, while “bravo”, “charlie” and “uniform” are long, distinct, ordinary English.

Digits are said either way — “seven” or “7”. The keys that are neither a letter nor a digit have one word each: space, dash, underscore, dot, delete, clear, done and cancel.

It is not a mode. Every utterance is tried as spelling first. If every word is a key, every key is pressed in order. If any word is not, nothing is pressed at all — the state line names the word it could not take, and the whole utterance lands in the field as a value instead. So “Shinrarta Dezhra” arrives whole, and “Alpha Centauri” fails on its second word and arrives whole too. Numbers are the case for spelling, and long names are the case for saying them.

A value that is itself one word of the alphabet — a system called Delta — is spelled. There is no escape word: say delta echo lima tango alpha and it comes out right.

This is live only while a keyboard is drawn. In the headset that is whenever the board is up. On the panel it is once the keys are showing, which is where a hearing that failed leaves you. Otherwise “bravo” in conversation reaches Directive 47 as the word it is.

Presses go to the drawn keyboard and nowhere else. Nothing here reaches the keyboard Elite is listening to.

Ask if you forget:

“how do I spell something” “the phonetic alphabet” “what is the word for K”

K is kilo.

Those are answered from the same table the keyboard parses, so Directive 47 cannot teach you a word the keyboard would refuse.

It knows what things are called

Every utterance is transcribed knowing the names around you: the system you are in, the station, the body, your next jump, your ship and its type, your carrier, the route ahead, and your fleet.

This matters more than it sounds. Proper nouns are where speech recognition fails hardest and least visibly. A misheard system name does not come back as an error — it comes back as a plausible English phrase. “Shinrarta Dezhra” becomes “shin arta desha”, everything proceeds confidently about the wrong system, and nothing anywhere reports a problem.

The names come from your journal. Nothing is looked up.

What happens to what you say

Spoken and typed questions run exactly the same path, so “where am I” means the same thing however you said it.

Nothing captured is written to disk or sent anywhere. Audio sits in a small buffer and is overwritten within about half a second unless the gate is open — because you are holding the key, or because Directive 47 heard somebody start talking. Only that stretch goes any further, and “further” means a speech model on your own machine.

A few small kindnesses:

• A press shorter than a quarter of a second is treated as a mis-press. Transcribing 80 milliseconds of room tone produces a confident wrong word, which is worse than nothing.
• If your microphone is unplugged mid-sentence, what was captured is dropped rather than transcribed — half a sentence you did not finish is worse than none.
• A key stuck down stops at 60 seconds and keeps the audio. You said something; better late than lost.
• Hands free, a stretch that turns out to have been somebody else talking is dropped without reaching the transcript, the panel or the log.

{"type":"object","properties":{},"required":[],"additionalProperties":false}

{"type":"object","properties":{"letter":{"type":"string","description":"One letter to give the word for. Pass all for the whole alphabet. Left out, D47 says how spelling by voice works."}},"required":[],"additionalProperties":false}