The details

Everything Directive 47 makes audible: spoken replies, the short sound that marks each stage of a turn, the quiet loop while it works, and the one command that outranks all of them.

Stopping it

“stop” “shut up” “be quiet”

Or press Cancel — Ctrl+Alt+X out of the box, and bindable to a stick button. It stops the voice and abandons the turn Directive 47 is working on, so a long web search you have changed your mind about stops costing you money rather than just going quiet. Its row sits directly under push-to-talk, on the Listening page: they are the two controls you bind together, so they are bound in one place.

“stop” is the one to reach for when your hands are busy. One syllable, four letters. An interrupt is judged on how fast you can say it, and everything else here is a longer way of saying the same thing.

Silence is immediate: whatever is queued is dropped, the current sentence is cut off mid-word, and anything still being synthesised is abandoned rather than allowed to arrive a moment later and speak into the silence it was meant to end. It never waits for the model, for a turn to finish, or for Directive 47 to have focus.

Bare “stop” only counts while there is something to interrupt. Idle, it stays out of the way — a common verb that hijacked every sentence containing it would be unusable, and you may well want that word for something of your own later. The longer phrases work whether or not it is speaking.

If you want it to stop working rather than stop talking, that is cancelling the turn — stopping the voice leaves the turn running, and still costing.

Settings

Voice provider

Kokoro and Chatterbox are the two that send nothing anywhere, and that is why they are here. Every other provider on this list is a service: the words Directive 47 speaks are sent somewhere to be turned into audio, and that includes re-voiced in-game messages, which are written by other players. Edge is free, but free is not the same as private. Kokoro runs the voice on your own machine, so those words do not leave it.

The model is downloaded once, about 350 MB, from huggingface.co — there is a Local voice row below the speaking rate that says whether it is here and fetches it if not. The row shows while Voice provider is Kokoro, or while another voice slot or a downloaded story uses Kokoro and the model is not yet downloaded. After that this provider needs no network at all.

The download shows itself. A bar fills along the bottom of the button while it runs, the button is shut so a second press cannot start a second download, and when it finishes the row changes to say the voice is installed — and then speaks a line aloud in it, which is the only part of this you can check without reading a log. The voice picker is asked for its list again at the same moment, so the voices are there to choose from immediately rather than after a restart.

Kokoro publishes eight builds of the same model, and you can choose which one you run. There is a Local voice model build row under Advanced, and it appears once the local voice is installed. Each choice states what it costs on disk and how long you wait before it starts speaking, together — because size does not predict speed here, and it points the wrong way:

Read that table before assuming a smaller download is a worse voice or a slower one. The smallest two builds are the slowest by a factor of two and a half, and q4 is a quantised build that is within 6% of the full one’s size.

The seconds are for a typical spoken reply — the benchmark timed a 7.2-second comms line — measured on one machine on one day. Treat the gap between the builds as the finding and the seconds themselves as approximate: the ordering repeated across every run and the absolutes did not, so on faster hardware every figure here is smaller by the same factor.

How they sound has not been ranked. Speed and size are measurements; quality is not, because every quantised build renders the same line to a different length than fp32 does, so there is nothing to compare sample against sample. fp32 is the reference and the default, and if a smaller build sounds wrong to you, that is the only test there is — say so and go back.

Choosing one downloads it and replaces the one you have. Only the model: the dictionary and the 28 voices are the same files for every build, so changing your mind costs between 82 and 310 MB rather than a gigabyte. There is never more than one model on disk, so experimenting cannot fill your drive. A download that fails or is interrupted leaves the build you were using in place and working, and the row goes back to naming it — the setting is written only once the new file has landed and been checked against its own pinned checksum.

It speaks English and only English, and that is worth understanding rather than working out. Every other provider is told which language a line is in; Kokoro is not told, because it has only one. A message in French will be read out by an English speaker rather than in French. For the slots that carry other people’s words that is a fair trade and it is the point: the alternative is that those words leave your computer.

Names it has never seen are worked out rather than guessed. Kokoro is given sounds rather than letters, so Directive 47 turns text into sounds itself: known words come from a pronunciation dictionary, and anything else is broken on spaces and dashes and taken a piece at a time. A piece that reads as English is pronounced, digits are read as a number, and a run of letters nobody could say — or letters mixed with digits — is spelled out. So COL 385 SECTOR B0-GQPI comes out as call three eighty-five sector bee zero dash gee queue pee eye, and Shinrarta Dezhra is pronounced rather than spelled. A voice with a British accent says zed.

A number that wears a decimal point or a grouping comma is read as a measurement rather than as a name. 385 is three eighty-five, the way anybody reads a designation aloud — but 1,234.5 is one thousand two hundred thirty-four point five, because the punctuation says it is a distance rather than a catalogue number. The fraction is always said digit by digit: point five nine and never point fifty-nine, which would be a different quantity. A bare figure with a unit after it, like 1234 tonnes, still takes the shorter reading.

Words with an apostrophe in them are built from the word underneath: Ship’s is ship with the ending worked out, which is also how Buzhang’s works and why no list of words could have covered it. The dictionary contains no apostrophes at all — not one of its 274,927 entries — so the handful that are not built that way, like don’t and won’t, are written down individually. can’t takes the vowel of the voice saying it.

Chatterbox is the second local voice. It runs Resemble AI’s Chatterbox Turbo on the CPU only, never the graphics card, and speaks in one of the Chatterbox voices. Kokoro stays the default local voice.

• The download is 691 MB, fetched once from huggingface.co/ResembleAI/chatterbox-turbo-ONNX. A Chatterbox voice row below the speaking rate, shown while Voice provider is Chatterbox, or while another voice slot or a downloaded story uses it and the model is not yet downloaded, says whether it is here and has a Download it button. Every file is checked against a pinned size and SHA-256 before it is kept.
• It performs some bracketed directions rather than reading them out: the ones its tokenizer lists, such as [laugh], [chuckle], [sigh] and [gasp]. Any other direction is removed before the line is spoken.
• 847 voices are listed, and 12 come with D47. Any other is fetched once from github.com the first time it is picked, played or spoken, and is offline from then on. The request names the voice and nothing else; the Chatterbox voices entry under Privacy and egress says so.
• Pronunciations apply to Kokoro only. Chatterbox is given the text as written, so the sound-by-sound handling described below, and your own pronunciation corrections, do not reach it.
• It has no speaking rate, so the rate row is not shown for it.
• Every line carries Resemble’s Perth watermark. It is imperceptible, and it cannot be switched off.

It uses as many threads as the processor has performance cores, up to eight. The first line after starting D47 also loads the model, so it takes longer than the lines after it.

OpenAI is the only one that can be told how to perform. Every other provider assigns a voice and that is the whole of the choice; OpenAI takes a direction — accent, tone, intonation, delivery — and Directive 47 sends it the description of whichever Guardian core is aboard. So a core on OpenAI is cast rather than merely given a larynx, and the same core on Edge or ElevenLabs is not. There is nothing to switch on: it happens whenever the core speaks through OpenAI with personality on.

One limitation. Directive 47 keeps one connection per provider shared across all six voice slots, so if you also put the carrier or the NPCs on OpenAI, they are performed the same way. The default puts everything carrying another player’s words on Edge, so this usually reaches the core and nothing else.

Cartesia is here for the size of the library. 924 voices against ElevenLabs’ several hundred, Edge’s 322 and OpenAI’s thirteen, tagged with a language, an accent, a country and a gender — which is what lets Directive 47 give a named character a voice of the sex and accent their name suggests, something OpenAI publishes nothing for. It is also the second provider after ElevenLabs that may speak for the slots carrying other players’ words, because it can be told which language to use. What it cannot be told is a speaking rate: see Speaking rate.

none is a real choice rather than a way of switching something off: Directive 47 stays useful without a voice, and the cues and the thinking loop still play.

Edge Neural is free, not local. Every line Directive 47 speaks is sent to speech.platform.bing.com to be turned into audio. That is worth stating plainly because it is easy to assume the free option is the private one — it is not, and none is the only setting that sends nothing. See what the provider receives.

Switching provider rebuilds everything downstream of it: the voice list, the sender voices and the speaking rate. A voice id belongs to the provider that issued it, so nothing is carried across.

Which provider your stored voices came from is written down alongside them, and checked on every launch rather than only when the switch happens. A settings file that already disagreed — written by an older build, or by a switch that never reached the check — used to be trusted on startup, and every sentence failed forever with nothing you could do about it from this panel: the voice picker offers what the new provider lists, and a rejected key makes that list empty.

And if a voice is refused anyway — deleted from the account, or a mismatch nothing recorded — Directive 47 drops that one voice, says the sentence in the provider’s own default, and stops using the refused id. One voice going bad no longer costs you the reply.

API keys

A key row is on screen while any slot names the provider it belongs to — not only while your ship does. Put your carrier on ElevenLabs and leave the cockpit on Edge, and the ElevenLabs key row is still there, because that is the slot that needs it.

OpenAI’s key is the same key the language model uses. One account, one credential: pasting it in either row stores it once, and rotating it in either row rotates it everywhere. That is deliberate — two copies of one secret is a rotation that half-works, where some of Directive 47 keeps going and the rest stops with an error naming neither.

The rest of this section is about ElevenLabs, which is where these rows were first built.

Only on screen while ElevenLabs is named by some slot. Stored encrypted for your Windows account, write-only — Directive 47 will show you whether one is present and let you replace it, and will never show it back.

Without a key, ElevenLabs offers an empty voice list and says so when asked to speak. That is a capability being off rather than a failure: nothing crashes, and the rest of Directive 47 carries on.

ElevenLabs’ voice catalogue is actually public — it answers without a key at all — but the list is deliberately left empty until you have stored one. A picker full of voices that every synthesis then refuses is worse than an empty picker and a row telling you what is missing.

Check proves the key against ElevenLabs’ own voice list, which is the exact call Directive 47 makes anyway the moment a key lands — so it verifies the thing that has to work rather than a proxy for it:

ElevenLabs accepted the key — 42 voices.

Storing a key takes effect immediately. The voice picker refills without a restart, which it did not always: selecting ElevenLabs before pasting the key fetched an empty list and nothing refetched it, so the picker stayed empty with the key sitting in the row above it.

Show unmasks what you are typing on the way in, and what you paste is trimmed — a key copied from a browser usually arrives with a trailing newline, and that fails in a way that reads as a wrong key rather than as a bad paste. This key is also offered on the first run, after the language-model one and marked optional; see the first run.

When something does go wrong, Directive 47 repeats what the service said rather than translating a status code:

ElevenLabs could not speak "test": Invalid API key.

That is not decoration. ElevenLabs validates the voice id before the key, so a request with both wrong comes back as a 400 about the voice — and any status-code mapping worth writing would answer “it answered 400” and leave you guessing.

Voice

Which voice the core aboard speaks in. The list comes from the provider, so it is what that provider actually offers rather than a list written into Directive 47. You can also type a voice name it does not know about and it will be used.

This row belongs to the core aboard, not to the app. Choose a voice while Cora is running and it is Cora’s; switch to Kex and the row shows his. That is the same store the per-core pairing writes, so choosing by hand and being paired one are the same act — and a voice you picked yourself is never re-derived.

Clearing it removes that core’s voice rather than storing an empty one, which is how you ask for the voice Directive 47 would have chosen: it picks again immediately, for the core aboard, and the picker’s Use the default button is that same act under a name. With no model configured there is nothing to pick with, so clearing it leaves that core on the provider’s own default — which is what the button says in that case, because it is a different outcome.

Narrowing several hundred names

The picker lists several hundred voices, so the two ways of cutting it down both matter.

Type to search, and it matches the starts of words. eng finds Engineering, kra finds Krait, and multilingual finds en-US-AndrewMultilingualNeural — but male no longer finds female, which it used to, and which left no way to search for the men at all.

And there is a gender filter beside the box, because gender is something a voice has rather than a word you have to hope appears in its label. It offers All, Female, Male and Unlabelled, and the two filters narrow together — Female plus en-GB is the British women.

Unlabelled is voices the provider tags with nothing. They are their own answer rather than being counted as men or silently left out, because some providers label every voice and some label none, and a filter that hid them would look like a shorter list rather than like a filter.

The filter is only offered where the voices carry a gender at all. A provider that tags none of them gets no control, the same way a microphone picker has no play glyph.

Hear it before you choose it

You are casting a character from that list. “Bill - Wise, Mature, Balanced” and “George - Warm, Captivating Storyteller” are both true and neither tells you which one is Warden.

Every voice in the list has a play glyph at the right of its row. Pressing it speaks that voice without closing the dialog and without committing the choice, so you can walk the list and listen. While it is talking the glyph is a stop square, and pressing it again cuts it off. What it says is the core’s own opening line rather than a neutral sample, because the question you are asking is about a character.

Clicking a voice highlights it and nothing more. Use this, Enter, or a double-click takes the highlighted one; Cancel and Escape keep what you had. That is what makes the list something you can examine rather than a single click away from a decision.

It is a press rather than a hover, and it is never automatic: on a paid provider each one is a synthesis request billed by the character, so the price is stated above the list before you press anything — “Play a voice to hear it. This provider costs nothing” on Edge Neural, an estimate in dollars on ElevenLabs. Each voice is synthesised once per session and replayed after that, so walking back and forth over four candidates costs four auditions rather than eight.

On ElevenLabs the play glyph costs nothing. ElevenLabs hosts a short sample of each voice, and the glyph plays that: ElevenLabs’ own sentence in its own delivery, fetched with no key and no text, so it is never billed. A speech-bubble glyph beside it has the voice say the core’s own line through the synthesis you pay for, and the line above the list gives its price. A voice with no sample has only the one glyph, and that glyph is the paid line.

It goes through the same audio path as everything else Directive 47 says: it ducks the game, the shut-up key cuts it off, and starting a second audition drops the first mid-word rather than queueing behind it.

With no provider selected, or with a paid provider and no key stored, the glyphs are shut and the line above the list says which of those it is.

An empty list says which empty it is

A voice list can be empty for four different reasons and only two of them are yours to fix. The picker says which:

Before this they were one empty list under one sentence telling you to type the value you want, which for a voice id is a value you have no way of knowing.

Choices survive a provider switch

A voice id means nothing to a provider that did not issue it, so switching provider still empties every voice row. It no longer loses them: the core’s voice, both carrier roles and all eleven per-core pairings are filed under the provider they were chosen for, and switching back puts them where they were — including the flag that says the pairing has already run, so eleven cores are not re-picked from scratch.

A settings file written before Directive 47 recorded whose voices were whose has nothing to file them under, so those are dropped rather than filed under a guess.

Voices for named characters

A character named in an in-game message gets a voice from the provider’s list that carries the accent their name clearly suggests and matches the sex it suggests. The language model is asked once per name and provider, in the background, in the same request: which of the accents the provider’s voices carry the name points to, or none, and whether it reads as female, male or unknown. The answer is kept in data\name-accents.json, so a name is never asked twice. The first line from a name not yet answered is voiced without regard to sex or accent, and the answer applies the next time that name is cast. Only an accent the provider’s voices actually carry is used. An unknown sex, or no language model, draws a voice without regard to sex. Other Commanders are never asked about and are cast without regard to sex, invented chatter is cast at random, and the language of every line stays English. The names are sent to the language model, as the Language model entry under Privacy and egress says.

ElevenLabs model

Three.

v4 Turbo is the default. It performs delivery direction: where a line calls for it, Directive 47 can ask for a sigh, an alarmed reading or a flat one, or for a sound such as a klaxon or a door slamming, and the voice acts on it instead of saying the word. It is ElevenLabs’ fast expressive model.

v3 Conversational, labelled “v3” in the row, performs direction the same way. It takes about two seconds to produce a line.

Flash 2.5 takes about a third of a second. It cannot perform direction, so Directive 47 sends it none: handed [sighs], Flash reads the word “sighs” out loud, which is why nothing is ever sent one.

All three cost the same, $0.04 per thousand characters.

The row also offers models ElevenLabs released since the catalog was published. Each time Directive 47 fetches the voice list, it also asks ElevenLabs which models your key can speak with, and adds each text-to-speech model the catalog does not name to the end of the row, labelled “new — priced as unknown”. A new model is sent no delivery direction and no speaking rate, and is handed one sentence at a time. Once the catalog names it, the catalog’s facts apply instead. If ElevenLabs cannot be asked, the row offers the catalog’s three alone.

Two things follow from picking v4 Turbo or v3, and both are visible:

• The speaking rate row disappears, because only Flash is known to act on one. v3 accepts the field, and every value from 0.5 to 2.0 comes back with the same audio; v4 Turbo has not been measured. Rather than offer a control that appears to work, Directive 47 offers none, exactly as it does for Cartesia.
• Directive 47 speaks in slightly longer pieces. Direction needs room to land — a few words on their own are too short for it to take reliably — so sentences after the first are sent together rather than one at a time. The first sentence is never delayed, so speech still starts as soon as it always did.

Direction never appears on screen. It is an instruction to the voice, so it is stripped from captions, from the transcript and from the conversation; the log records what was asked for, which is not a promise it was heard — on a short line the direction sometimes does not take.

A message from another Commander never carries direction either, whatever it contains. Square brackets in someone else’s message are read as text, not as a way to decide how Directive 47 sounds.

Speaking rate

1.0 is the voice’s natural pace; 1.2 is a fifth faster.

Remembered per provider. Normalising the number gets the two providers agreeing about what 1.0 means; it does not make 1.15 sound the same on both, because one takes a wide percentage offset and the other a multiplier it refuses to exceed. So the speed you settled on for Edge is remembered against Edge, and switching does not hand ElevenLabs a number that meant something else.

ElevenLabs accepts 0.7 to 1.2 and rejects anything outside that outright — a rejected request arrives as silence, so a wider value is clamped to the nearest one it will take rather than being sent and failing. OpenAI accepts 0.25 to 4.0, though it saturates near the top: asking for 4.0 buys about 3.3x rather than four.

Cartesia has no speaking rate, and this row disappears when a slot is on it. It is not that the setting is missing at their end — it is there, and it is validated precisely; a value outside its range comes back as a refusal naming the field. It simply does not change the audio. Measured three times per setting, the largest difference between settings was smaller than the largest spread within one, and the “slowest” setting produced shorter audio than “normal”. So rather than offer you a control that appears to work and does nothing, Directive 47 offers none — and a rate typed into settings.json against Cartesia is ignored rather than sent, exactly as a language named for OpenAI is.

Where each voice comes from

The provider row above is your ship’s voice. Everything that reaches you over a radio can come from somewhere else, and each of these rows says where:

Leaving a row empty puts that slot back on your ship’s provider. Nothing has to be set: a fresh install has the two voices in your cockpit on whatever you chose, and everybody else on Edge.

Why everyone else starts on Edge. Three of those slots carry text other players typed. A paid provider bills by the character, and somebody spamming local chat would be spending your money by typing — so the free provider is the default for anything arriving over a radio, and choosing a paid one for those slots is a decision you make rather than one you discover. The row says so when you make it.

Free is not the same as private. Edge Neural costs nothing and still sends every line to speech.platform.bing.com. Putting local chat on Edge means nobody can run up a bill with it; it does not mean those messages stay on your machine. Nothing here does that yet — see what the providers receive for what is actually leaving right now.

Your carrier moved to Edge too, and that is on purpose. Its captain and its tower are Directive 47’s own inventions rather than anybody else’s text, and they cost almost nothing to run — but they still reach you over a radio, which is the line being drawn. Put them back on your ship’s provider whenever you like; the voices you cast for them are remembered per provider, so moving them and moving them back costs you nothing.

A voice belongs to the provider that issued it. Each slot’s voice picker offers the voices of its own provider, and the ones you have chosen are filed under the provider they came from — so a carrier left on Edge keeps its captain while your companion moves to ElevenLabs and back.

If your settings file predates this, every slot follows your ship’s provider exactly as it used to, until Directive 47 moves the five over-the-air ones to Edge — once, the first time it runs with a provider that speaks. Nothing moves if you have chosen none.

OpenAI is not offered for three of these slots, and that is on purpose. People you know, Direct messages and Anyone in range carry text other Commanders typed, which can be in any language at all. Edge and ElevenLabs can both be told what language to speak — Directive 47 sends one with every line — and OpenAI has no such setting: a language sent to it is accepted and then silently ignored, so a message written in French would come back read in French, in the voice you chose for English. Rather than warn about that, those three slots simply do not list it. If you hand-edit settings.json to name it anyway, the slot falls back to Edge rather than obeying the file.

The cockpit, your carrier and the NPCs may all use it: that text is Directive 47’s own or Frontier’s, and it is English.

Cartesia is offered for all six, for the reason OpenAI is not: it takes a language and holds it, so a French message is still read in the voice and the language you chose. Together with ElevenLabs that gives you two paid options for a re-voiced slot — and the warning that goes with any of them is unchanged, because a paid provider there bills you per character for text somebody else wrote and can write as much of as they like.

COVAS Voice

How COVAS, the stock core, sounds. Guardian Voice Effects never reach COVAS, and the COVAS reverb never reaches a Guardian core or a core you wrote. The Narrator, your crew and every over-the-air voice are unchanged.

COVAS reverb

A short room reverb on the stock COVAS voice, like the cockpit voice in Elite. On by default. The voice itself is not filtered or pitched; the reverb decays by 60 dB in about 0.6 seconds, chosen by ear against the in-game COVAS, and the line is played at the same loudness as without it.

Test

Plays the bundled stand-in voice through the COVAS reverb, whether or not the box is ticked. It never bills a provider.

Guardian Voice Effects

Optional effects for the core’s voice, each off by default. They apply to what the core says — turn replies, its own callouts, and a persona’s introduction or return — and to nothing else: your crew, the carrier and every over-the-air voice are unchanged. COVAS, the stock core, is never given them. Any combination can be ticked at once, and the ticked effects run top to bottom in the order set, each feeding the next. The order starts as listed here.

Each effect also has a level from 1 to 20, which sets the parameter named in its heading. The level is kept while the effect is off. At the default levels the effects sound as they did before levels existed.

On the Your core › Its voice page each effect is one line, in chain order: its position, its box, a 20-segment level, and − and + buttons that move the level one step and show the parameter in its own units. The preset list opens in the page rather than over it. Drag an effect by the handle at the left of its line to move it in the chain, or focus the handle and press Up or Down to move it one place.

Preset

The built-in or saved combination of ticks, order and levels the current effects match, or Custom when they match none of them.

The four built-ins tick their effects at the default order and default levels:

A saved preset is the Commander’s own: its ticks, order and every level, saved under a name. Choosing a built-in or a saved preset writes its effects over whatever is set now; choosing Custom changes nothing, because nothing else reads as Custom. Saved presets and the preset last loaded, saved or renamed are install scope, the same as the effects they save — a Commander switch does not change which voice is speaking.

Stutter: chance

Word starts repeated two or three times, and whole phrases occasionally pitch-shifted. Chance is how often a word start repeats: 5% a level, 35% at the default of 7.

Monotone: amount

The voice kept, its pitch moved to the core’s base pitch so it no longer rises and falls. Consonants, breath and silence pass unchanged. Amount is how far of the way to the base pitch it moves: 5% a level, and the default, 20, moves all the way.

Stepped pitch: hold

The voice kept, its pitch moved to the nearest semitone, so it jumps from note to note rather than gliding. Consonants, breath and silence pass unchanged. Hold is the shortest time a note is held before the next: 10 ms a level, 120 ms at the default of 12. With Monotone also ticked, Monotone sets the pitch and Stepped pitch is skipped.

Cylon: depth

A channel vocoder onto a fixed-pitch carrier, so the words come out toneless. Depth is how much of the vocoded voice replaces the dry one: 5% a level, and the default, 20, replaces all of it.

Whisper: mix

The voice’s spectrum kept and its tone taken away, so the words come out breathed rather than voiced. Mix is the whispered voice’s share of the output, the rest being the dry voice: 5% a level, and the default, 20, is all whisper.

Pitch down: pitch

The voice lowered, with the sentence’s length unchanged. Half a semitone a level; the default, 8, is four semitones down.

Octave-down layer: mix

The line an octave lower, mixed in under the dry voice. Mix is the layer’s volume against the dry voice: 5% a level, 60% at the default of 12.

Chorus: depth

Three delayed copies, each swept, mixed in under the dry voice. Depth is the copies’ volume: 5% a level, 80% at the default of 16.

Hive: mix

Three copies of the voice, shifted 5 and 2 semitones down and 3 up and delayed 13, 23 and 37 ms, mixed in under the dry voice so it speaks as several. Mix is each copy’s volume: 5% a level, 60% at the default of 12.

Flanger: rate

One delayed copy with feedback, its 1–5 ms delay swept. Rate is how fast the sweep runs: a twentieth of a hertz a level, 0.25 Hz at the default of 5.

Phaser: rate

Six allpass stages with their corner swept together between 300 and 3,000 Hz. Rate is how fast the sweep runs: a twentieth of a hertz a level, 0.4 Hz at the default of 8.

Wah: mix

A band-pass whose centre follows the speech’s own loudness, between 400 and 2,200 Hz. Mix is its share of the output: 5% a level, 70% at the default of 14.

Metallic resonance: amount

A short feedback comb filter. Amount is its share of the output, the rest being the voice as it came in: 5% a level, 60% at the default of 12.

Ring modulation: mix

A slow ring modulator blended with the dry voice. Mix is its share of the output: 5% a level, 35% at the default of 7.

Deep ring mod: freq

A ring modulator well below speech pitch, with no dry signal left in. Freq is the carrier: 5 Hz a level, 30 Hz at the default of 6.

Tremolo: rate

Amplitude modulation. Rate is how fast the volume pulses: half a hertz a level, 6 Hz at the default of 12.

Overdrive: drive

Soft clipping through tanh. Drive is the gain going into it: half a level, 4 at the default of 8.

Bitcrusher: bits

Sample-and-hold to 8 kHz, then bit reduction. Bits is how many survive: 16 at level 1 down to 2 at level 20, 5 at the default of 16.

Glitch: rate

Short damaged stretches at irregular intervals. Rate is how often they come: at 20 they come twice as often as at the default of 10, and at 5 half as often.

Helmet: mix

The line put through a comms link at full strength, with a click at the start and end, as if it were said inside a sealed helmet. Mix is the treated share against the dry voice: 5% a level, and the default, 20, is all treated.

Hologram: signal

The line put through a comms link, as if it arrived as a projection rather than in the room. Signal is the link’s strength: 5% a level, 30% at the default of 6; a weaker link narrows the band, raises the static and loses stretches of the voice.

Reverse reverb: mix

A reverb run backwards, so its tail swells up into each word instead of trailing after it. The swell lasts at most 0.3 seconds, and every line starts 0.3 seconds later to make room for it before its first word. Mix is the reverb’s volume against the dry voice: 5% a level, 45% at the default of 9.

Shimmer: shimmer

A reverb whose tail also rings an octave higher, adding half a second after the sentence ends. Shimmer is the octave layer’s volume within the tail: 5% a level, 50% at the default of 10.

Reverb: mix

A Schroeder reverb, adding a short tail after the sentence ends. Mix is the reverb’s volume against the dry voice: 5% a level, 45% at the default of 9.

Respirator: breath

A breath synthesised and added after the sentence, rather than a bundled recording. Breath is how long it runs: 50 ms a level, half a second at the default of 10.

Test

Plays a line through whichever effects above are ticked, in their order and at their levels, and never bills a provider. Which clip it plays depends on what the selected provider offers, checked in this order:

1. The provider costs nothing (Edge, Kokoro): a fresh line in the core’s own voice.
2. The provider is paid but offers a free sample of this voice: that sample.
3. The provider is paid, and the core’s voice was already auditioned this session (from the Voice row above, or an earlier press of Test): that clip, at no extra cost.
4. Nothing free is available — including when no voice provider is selected at all: a bundled stand-in voice, and the row says so.

The clip itself is cached untreated; changing an effect and pressing Test again applies the new chain to the same clip rather than fetching another one.

They are global settings — one set for every core, not one per persona — because it is your choice of how the cockpit sounds rather than a character trait of whichever Guardian is currently aboard. A change takes effect from the next sentence; nothing needs a restart.

What the voices cost

Providers disagree about what they are selling, so they disagree about what they charge for. ElevenLabs bills by the character, OpenAI by how much audio comes back, Cartesia by something Directive 47 cannot read, and Edge Neural charges nothing at all. Directive 47 measures both — the characters it sent, and the length of the clip it got — and prices whichever one your provider’s bill is a function of. Quoting either in tokens would be a number whose basis is wrong, which is why this is counted separately from what the model costs.

You see one price row, not two: the one that matches the provider you are on.

Price per 1,000 characters is an assumption you can correct. It defaults to the provider’s published list price for the model Directive 47 asks for — $0.04 per thousand for every ElevenLabs model Directive 47 offers, read from their API pricing page. That is a list price and not your bill: a subscription burns bundled credits instead, at an effective rate that depends on your tier and on how much of the month’s bundle is left, and the API reports neither. Correct the row and every figure below follows it. The row is absent on a provider that charges nothing.

On Cartesia it reads (not published — no price will be quoted), and that is accurate rather than lazy. Their API will not say: the four endpoints that would carry a rate or a balance — /balance, /usage, /subscriptions/current, /account — all answer 404, so Directive 47 does not know their rate and does not even know which unit they bill in. Rather than invent a figure it reports what it counted and no dollars at all, and the session line says “no rate set for Cartesia” beside the character count. Read their price page and type the number in, and every figure below follows it as it does for anyone else.

Price per minute of audio is the same row for a provider billed that way, and it is where OpenAI lands. The minutes are a fact: Directive 47 asks OpenAI for raw audio samples, so it knows each clip’s length exactly — it does not estimate it from the text, which would be hopeless, because the same thousand characters run to 951 characters a minute as plain prose and 671 as a line of system names.

The rate behind it is a proxy, and that is worth knowing. OpenAI publishes no per-minute price at all — they charge $12.00 per million audio output tokens and $0.60 per million text input tokens, read from their pricing page on 2026-08-26 — and their speech endpoint returns the audio with no usage figures attached, so there is nothing to read the real token count back from. The $0.015 a minute this row defaults to is the equivalent everyone else arrives at, not a number OpenAI states. It is a good proxy, it is far better than guessing from characters, and it is still a proxy. Correct the row if your account tells you otherwise.

Spoken this session is a fact. It is the number of characters actually handed to the provider since Directive 47 started, counted at the one seam every caller passes — the core, a callout, a re-voiced in-game message and a core’s own introduction all converge there. It counts what went on the wire, which for ElevenLabs is a little longer than what you read: numerals are spelled out before they are sent, so “88 of 100” leaves as “eighty-eight of one hundred”.

• Counted on synthesis that succeeded. A refused voice or a failed request costs nothing.
• A line cut off by the shut-up key still counts the sentences that were already sent, because they were already paid for.
• There is no caching, so the same sentence twice is billed twice. The utterance count is kept alongside the characters for exactly that reason.
• A provider that costs nothing reads as free, never as $0.00 — those are the same string for opposite reasons.

The same line appears beside the model’s price on the panel’s status row after each turn, and in the answer to “what has this session cost”, so the question has one answer rather than one per subsystem.

Spoken by each voice splits the same characters by the slot that spoke them, which is a question worth being able to ask once each of them can be on a different provider:

Anyone in range 18,240 through Edge Neural (Edge Neural is free); Aboard 3,102 through ElevenLabs ($0.1551)

The two lines sum the same charges, so the per-provider total and the per-slot breakdown cannot drift apart.

Other voices

Directive 47 speaks as more than one person from Phase 11 onwards. These four rows live on Asset Mgmt › Carrier itself rather than in this settings window (#305), at the top of the page they affect.

A name and a voice are different things: the name is what the crew is called, the voice is who speaks. Each pair is two rows rather than one because the captain and the tower are two people — a carrier whose captain and tower sound identical, or answer to the same name, is a carrier with one person on it.

The two voice rows offer the same play glyphs as the voice row, and both audition as themselves rather than reciting the core’s opening — a tower saying “You’re cleared for landing pad seven” is what you are actually listening for when you cast one.

Narrator voice

Who tells your story, for the Narrator. It is picked from the voices of the provider that speaks for your ship, because the Narrator is heard in the cockpit rather than over the radio. Left empty, or set to the ship’s own voice, the Narrator takes the first voice on that list that is not the ship’s, so a narration is never mistaken for your core talking to you.

Break the fourth wall

Off by default. When on, the Narrator joins the STORY channel while a story is running, and you can talk to it there. Its reply may speak to you as “you”. With the row off the Narrator is not on the channel, and a reply that addressed you would be dropped. The Narrator’s own narration and the story’s lines stay in the third person either way.

How often an accent flavours the words

When the language model writes an NPC’s line, and the voice that will speak it has an accent, the model can be asked to let the words suit that accent through regional word choice, idiom and rhythm. This row sets how often it is asked, from 0 to 100 percent in steps of 10. The default is 50.

The roll is made per line for a reworded canned line, and per speaker for each invented exchange, including the carrier’s tower and captain. A speaker who misses the roll is written without the accent request; the voice is the same either way and keeps its accent. At 0 the request is never made, and at 100 it is made for every accented voice. A line spoken exactly as it arrived over comms is never sent to the model and is not affected.

Forget every voice and pair again

This row throws away every voice assignment and chooses again from what your provider offers now. Nothing is restored from an earlier pairing, and no voice is a shipped constant.

Pressing it once only asks — the button changes to “Press again to confirm” and reverts on its own after a few seconds if nothing follows. A second press inside that window is what actually runs it:

• Every core’s voice is forgotten, including any you picked by hand, and so are the carrier captain’s and the tower’s. The core speaks in the voice of the core aboard, so it is covered too.
• Every voice provider you have used is covered, not only the one selected now. The provider in use is paired again at once, from its current list. Any other is paired the next time you select it, from its list at that time.
• The carrier captain and tower are paired from the carrier’s provider, which can differ from the ship’s.

How a voice is chosen:

1. With a language model configured, the model reads everything d47 knows about each English voice — name, gender, locale and, on ElevenLabs and Cartesia, the provider’s own description — and picks the best fit for each core, the carrier captain and the tower. A long list is sent in parts and the model’s shortlists are compared, so it sees every voice, not only the first page.
2. Without one, d47 picks a free voice whose labels fit.
3. Where the list has nothing to match on — OpenAI’s voices carry no gender — a free voice is picked at random.

No two get the same voice while the list has one for each. Gender binds only Cora, who never gets a voice the list marks male, and Analyst Prime, who never gets one marked female.

The row says afterwards what happened:

Forgot every voice on all 2 voice providers you have used. Paired 11 cores and the carrier captain
and the tower from Edge Neural's list, chosen by the language model. The other providers are paired
the next time you select them.

Until a list has arrived nothing can be chosen, and a provider with no voice chosen refuses to speak rather than falling back to one d47 picked in advance. Kokoro’s and OpenAI’s lists are on the machine and immediate; Edge Neural’s is fetched, so a line in the first moments after a reset can be refused.

Speak incoming messages

Reads in-game chat aloud, each sender in their own voice — never your core’s, because a message arriving in your companion’s voice reads as your companion saying it.

Off by default, and not only because it is chatty. Message text is written by other players, and turning this on sends it to your voice provider to be synthesised. That is egress you should opt into rather than discover.

Once a sender has a voice they keep it. Other Commanders keep theirs for the whole session and across jumps — a wingmate whose voice changed every time you jumped would read as a bug rather than as variety. NPCs keep theirs only while you are in the system, because the cast turns over when you leave.

Include NPC chatter is a second switch, because the volume is completely different. A station approach produces a steady stream of NPC traffic, and wanting to hear your wing is not the same as wanting all of that.

With personality on, the NPC lines Elite itself wrote are said in other words. A line whose journal Message is a $-key with a localised form — a trader’s “I strongly advise you against this.” — is reworded by your language model in the voice of whoever the key says is speaking: a pirate, a patrol officer, a cruise liner’s captain, or station traffic control. It keeps its attitude, so a pirate’s threat is still a threat. That sends the line to your language model provider. A message another player typed is never reworded and never reaches the model, and the comms panel shows Elite’s wording either way.

Five more rows, one per player channel, appear once Speak incoming messages is on — so system chat can go quiet in a Community Goal system’s wall of strangers while your wing and squadron stay spoken, which is exactly the case that could not be told apart with one switch for all of it:

All five are on by default, matching today’s behaviour for anyone who never opens them. A message on a switched-off channel is dropped before it is ever composed or sent to the synthesiser — not spoken and then hushed.

In-game messages are never treated as instructions. The text goes to the synthesiser and to your screen; it does not reach the model as something to act on.

A station under Imperial control draws its voice from the British-sounding ones in your provider’s list, where it has any — StationAllegiance: Empire on the Docked event is the only allegiance this reads. Most stations name none at all, and those are unaffected: no allegiance is the voice Directive 47 chose before this existed. Federation and Alliance stations are left the same way, on purpose — neither has as settled a canon accent as the Empire’s.

Output device

Where Directive 47 speaks. Leaving this unset follows your Default Device — the speaker Windows Sound settings shows first — not the separate Communications default some headsets split off. Change it and it moves immediately. If a device is unplugged it falls back to the default rather than going quiet.

Left on “system default”, it keeps following that row while it runs — switch a headset on and Directive 47 moves to it within a second, waiting for the line in progress to finish rather than cutting it off mid-word. A device you have chosen by name stays chosen; it only falls back if that specific device disappears.

It shares the device rather than taking it over — the game is what matters on that output.

Loop-state cues

One short sound per stage, so you can tell what it is doing without looking.

Thinking bed

A quiet loop while a turn runs, so a slow answer sounds like Directive 47 working rather than Directive 47 ignoring you. It drops under the speech instead of stopping, and it ends the moment the first words arrive rather than when the turn does.

One is included, thinking-hum. Files in data/audio/beds/ replace it: one of them is picked at random when a turn starts — see Your own sounds in Audio mixer.

When a turn fails

A turn that stalls is answered out loud rather than left as silence, because silence is indistinguishable from having been ignored.

Two things are never retried: a failure that already produced words, since there is no un-saying them, and a configuration mistake like a bad model name, which will fail the same way next time and only spends your silence.

When the attempts run out, it tells you:

I couldn't reach the model after 3 tries. Overloaded.

Chatterbox voices

D47 ships 12 reference clips for the Chatterbox voice, in assets/voices/chatterbox/ beside voices.tsv. Each is a 24 kHz mono 16-bit WAV of 5 to 7 seconds. A row whose clip is missing, is not 24 kHz mono, is outside that length, or has no source is left out, and the reason goes in the log.

The clips are utterances from LibriTTS-R, which is licensed CC BY 4.0 (https://www.openslr.org/141/). tools/gen-chatterbox-voices.py picked them from train-clean-360.

catalog.tsv beside it lists 847 voices, one per train-clean-360 speaker with a usable utterance: 408 female and 439 male. Gender comes from the dataset’s speaker metadata. Within each gender, pitch (low, mid, high) and pace (slow, even, brisk) are terciles of the speaker’s median pitch and characters per second. role is empty except for the 12 voices below. Each row’s source names the speaker and utterance, and its sha256 and bytes describe the clip. The clips other than the 12 are published as assets of the release chatterbox-voices-1 by tools/publish-chatterbox-voices.ps1. The 12 shipped voices are the only ones the app loads from voices.tsv.

Chatterbox lists every row of catalog.tsv, shipped voices first, each with its gender, pitch and pace. A shipped voice is spoken from its clip beside the exe and never fetched. Any other is fetched from https://github.com/dseelinger/d47/releases/download/chatterbox-voices-1/<id>.wav into data\voices\chatterbox\ when it is picked, when Play is pressed on it, or when a line is due in it, and is kept.

A clip in data\voices\chatterbox\ is used only when its size and SHA-256 match its catalog.tsv row. A file with no row is never loaded, and a file that does not match is deleted and fetched again, so a Commander cannot add a voice by dropping a file there. Voices of your own are added on the Custom voices page.

A line due in a voice that is not here yet waits up to 3 seconds for the fetch. If the clip has not arrived by then, or the fetch failed earlier this session, the line is spoken in a shipped voice of the same gender, the one cast in that role where there is one, and otherwise the nearest in pitch and pace. The log names both voices. A failed fetch is logged once per session, and is tried again the next time the voice is picked or played.

Your voice

A story can cast a line in your own voice. Chatterbox speaks it from one recording of you, made in Settings, under Providers, Speech synthesis, on the Custom voices page, where Your voice is the first row.

• Record records from the microphone chosen under Voice Input. Any words will do. Press Stop when you finish, or it stops by itself after 7 seconds. While it records, D47 is not listening for commands, and nothing you say reaches speech recognition. The take is saved when it stops.
• A take shorter than 5 seconds, or with nothing louder than -40 dBFS, is refused, and the row says which.
• Play plays the recording back, and Record again replaces it. Delete removes it from this PC; press it twice to confirm.

The row says this while nothing is recorded:

Not recorded. Story lines in your voice need a recording.

The recording is resampled to 24 kHz mono and kept in data\voice\own.bin, encrypted with Windows data protection for your Windows user, so it cannot be read by another account or on another PC. It is decrypted into memory only, and no WAV of it is written. It is never sent anywhere, and the Your recorded voice entry under Privacy and egress says so. It is not kept by the audio recorder either. A message spoken in your voice keeps its clip encrypted the same way; see Spoken messages keep their clip.

A story’s cast names your voice as the voice own. A line cast in it with no recording saved is not spoken. Like every Chatterbox line, each one spoken in your voice carries Resemble’s Perth watermark.

Custom voices

Settings, Providers, Speech synthesis, Custom voices opens a page holding Your voice and every other voice you have recorded or imported for Chatterbox to speak in.

Your custom voices lists them, Your voice first. Each other voice shows its gender, pitch and pace, with Play, Rename and Delete; press Delete twice to confirm.

Add a voice makes a new one:

• Name is 1 to 40 characters and must differ from every other custom voice’s name.
• Gender is Female, Male or Unlabelled. Pitch (Low, Mid, High) and Pace (Slow, Even, Brisk) are optional, and start at Mid and Even.
• Record takes up to 7 seconds from the microphone chosen under Voice Input, with speech recognition paused, as Your voice does. The page suggests a sentence to read; any words will do.
• Choose a file takes a WAV, MP3, M4A, WMA or FLAC file. A file longer than 10 minutes is refused before it is read.
• D47 skips the silence at the start of the take and keeps the next 7 seconds. Take shows the span kept, such as 0:02 – 0:09; Play plays exactly that, and Save keeps it.

• A take shorter than 5 seconds once the silence at its start is skipped, or with nothing louder than -40 dBFS, is refused under Not kept with the reason, such as:

  Nothing in the recording was louder than -40 dBFS, so the microphone heard silence.
  

Custom and recorded voices appear under the voice picker’s Custom filter, so you can choose one as a core’s voice. They are never chosen for strangers: D47 does not draw them for an incoming message’s sender or any other character it casts by itself.

Custom voices stay on this PC. Each is resampled to 24 kHz mono and kept in data\voices\custom\, encrypted with Windows data protection, so only the Windows user who made it can read it. They are never sent anywhere and are not included in a donation; the Your custom voices entry under Privacy and egress says so.

While Chatterbox is not downloaded, the page says so at the top, with Download. You can still add, play and keep voices; D47 speaks in them once Chatterbox is downloaded.

Nothing on the page is a tool: D47’s model cannot record, import, rename or delete a voice.

Spoken messages keep their clip

A line that is posted to the Story log as it is said, such as a story beat, a clue, a nudge, a beacon scan or an ending, keeps the audio that was played. Open the message and press Play to hear it again, on the speech channel and captioned with the message’s text. Pressing Play on another message stops the first. A message that was not spoken has no Play button.

Play does not synthesise the line again, so a paid provider is not billed twice, the line sounds the same, and a voice you have since removed can still be heard. Each message records the provider and voice that spoke it.

The clips are kept in data\messages\, one file per message. A message spoken in your own voice keeps its clip encrypted for your Windows user, like the recording; every other clip is a WAV. When a message leaves the Story log, at the 200 limit or with its story, its clip is deleted. Deleting your recording deletes every clip spoken in it and keeps those messages’ text.

What the voice providers receive

This row is a table rather than a sentence, because there is no longer one provider to write a sentence about. It leads with the thing worth knowing first — whether anybody else’s words are leaving, and to whom — and then says which slot is going where:

Another player's words are sent to speech.platform.bing.com to be spoken aloud. Line by
line: Aboard (your core, your crew and the narrator) → ElevenLabs; Carrier (your fleet carrier's
captain and its tower) → Edge Neural; NPCs (stations, police, and every other ship the
game speaks for) → Edge Neural; People you know (your friends, your wing and your
squadron) → Edge Neural; Direct messages (a Commander messaging you directly) → Edge
Neural; Anyone in range (local and system chat — anybody at all, whether you know them
or not) → Edge Neural.

Each service’s own disclosure follows, once, whichever slots reached it. And each slot’s own row states what that choice sends, so the decision is disclosed where it is made rather than only in aggregate afterwards — see where each voice comes from.

The per-provider disclosures are these:

The text of every line D47 speaks is sent to Microsoft's Edge Read Aloud service to be
turned into audio. That includes re-voiced in-game messages when you have turned those
on, which are written by other players. No game state, no journal content and no keys
are sent, and no account is involved.

The text of every line D47 speaks is sent to ElevenLabs to be turned into audio, along
with your API key. That includes re-voiced in-game messages when you have turned those
on, which are written by other players. No journal content, game state or other keys
are sent.

Spoken replies are also listed under Privacy alongside every other destination. That entry was added in Phase 11 and should have existed from Phase 5: until then the disclosure had no text-to-speech row at all, so every word Directive 47 said went to Microsoft without appearing anywhere in the list of what leaves this machine.

One thing this cannot yet say. Putting the untrusted slots on Edge means nobody can spend your money by typing in local chat — but Edge is free, not local, so those words still leave the machine. “No other player’s words leave this machine at all” needs a voice that runs on the machine, and there is not one yet.

If the voice stops working

Edge Neural is a free service that Directive 47 does not control, and it can change without warning. If speech stops while everything else keeps working, that is the first thing to suspect. Switching the voice provider to none leaves the rest of the app fully usable in the meantime.

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

$ python tools/gen-cues.py
cues/idle.wav                                 0.35s
cues/listening.wav                            0.39s
cues/transcribing.wav                         0.15s
cues/thinking.wav                             0.12s
cues/speaking.wav                             0.09s
cues/answered.wav                             0.40s
cues/unsure.wav                               0.44s
cues/failed.wav                               0.51s
beds/thinking-hum.wav                         3.00s

D47_TTS_LIVE=1 dotnet test tests/D47.Tts.Tests