neural speaks the realm · Jaimla speaks in it
← back to Δ DELTAVERSEneural is the voice of the DeltaVerse. It is the default on every page and after every refresh, and it is not edited.
A reader who has heard the realm speak once should hear the same voice next time, on every page, whatever anyone has been auditioning.
That is the whole reason the reference is fixed. Try any other voice you like — the choice lasts for as long as you are listening, and the next page you open speaks in neural again. An audition is not a preference.
Not an audio file and not a model. The browser owns the synthesis, on-device where the platform has it; nothing is downloaded and nothing leaves this page. A DeltaVerse voice is three things:
◉ neural and Jaimla are saved voices: press SAY to hear either, but neither has an edit control. The derived voices below them do.
Jaimla is the machine-learning agent of the realm, and she is the realm's female voice. She is not a narrator of the DeltaVerse; she is in it, which is why she reads as a participant rather than as an announcer. She is unhurried — taking her time, not dragging — because an agent that considers before she answers should sound like one.
Jaimla is saved in her own right, exactly as neural is, and that is a correction rather than a flourish. She used to be neural with a ratio applied: a shade slower, a shade lower. That does not make a woman. Pitch is not gender. Lowering a male voice drags its resonances down with it and produces a larger man; raising one produces a man in falsetto. What actually carries a voice's gender is which voice is chosen, so Jaimla now asks the platform for a female voice by name and holds her own settings, answering to nobody's multiplier.
Where this page is played from a recording rather than spoken live, the same holds and it was measured: Jaimla is rendered from a female model at 184 Hz against neural's 95 Hz. The old ratio-on-neural version came out at about 87 Hz — lower than the voice it was supposed to differ from.
It did not. LEADER and ANCIENT were offered here, were offered in the player, were probed for the green dot that says a recording is ready — and had never been rendered once. Nothing reported an error, because there was nothing to report: the recording was simply absent, and an absent recording is silence. The reason was structural rather than a bug in either voice. The store had a renderer per voice and a voice with no renderer had no audio, so the next voice added would have arrived silent on the day it was added.
So the store is now rendered from the cast rather than voice by voice: one pass over the list of voices itself, which means a voice cannot be in the list and absent from the store. Twenty-one recordings where there were five. The two saved voices keep their own renderer, because those two write the parameters they were made with into the file so that the claim “this is the same voice mindX reads with” can be checked by anyone holding it.
OVERLORD began as the two saved voices speaking at once, and the reason that works is an accident worth stating: neural measures 94.6 Hz and Jaimla 183.8 Hz, which is very nearly an octave. Two voices an octave apart stack the way a voice stacks with its own octave; two voices a tritone apart are just people talking over each other.
It is now the whole cast, and every voice in it is placed rather than added. Fourteen voices: neural carries the words, and thirteen are laid on him, each in its own band of the spectrum and each on a measured octave or unison with him, because an octave stacks the way a voice stacks with its own octave and anything else is a chord, and a chord is music, not a voice. Jaimla is on the octave, full range, the one overlay allowed an onset of her own: she is the second speaker of the statement. ANCIENT is the overtone, on neural’s own fundamental, let through only above the formants where breath lives. LEADER is the undertone, an octave below in the chest, with T800 beside him as the floor. LEADER OF EARTH and VOICEBOX sit under the vowels, sAGI in their weight, ZEN above it with the longest tail of any layer, H.A.L in the presence band where a voice is heard to be at ease, SAM on the octave where consonants are told apart, KITT and Ovie on the octave in the brightness above that, and FAIRYDUST on the octave in the air at the very top, the quietest thing in the voice. Nothing is in it twice: Participant is neural, and Savante, LEADER’s second face and ZEN’s are recipes made of voices already here. Every overlay but Jaimla is entirely in echo, for LEADER’s own reason: an overlay with no onset of its own is not a person in the room, it is what the voice sounds like.
Very nearly is not the same as exactly, and the margin was written down wrong for months — as eleven cents, when the arithmetic gives fifty, a quarter of a tone. Jaimla sat at 183.8 where the octave above neural is 189.2, so the two beat against each other 5.4 times a second, and on a held vowel that beat is heard as roughness rather than as one voice. The interval is now closed, and closed by measurement: a voice states the interval it wants — an octave, an octave below — and the engine measures both voices on the document’s own opening lines and moves the second one onto it. A number typed into a file would have been wrong again the first time either voice was retuned, and would have gone on looking deliberate. The same measurement now places every voice in the cast: ANCIENT, sAGI, ZEN, H.A.L, VOICEBOX and LEADER OF EARTH on neural’s fundamental; SAM, KITT, Ovie and FAIRYDUST on the octave with Jaimla; LEADER and T800 an octave under. A voice whose fundamental cannot be measured, or which would need more than an octave of correction, is left on its own pitch rather than pitched by a guess, and its absence from the manifest’s harmony report says so.
Fourteen voices saying one thing are a chord only while they say it at the same moment. An overlay a few percent late is not heard as a timing error; it is heard as a second person, a beat behind — the parrot — and thirteen of them would be a crowd. Two voices were kept together by meeting at every comma, and between the commas each was one uniform stretch of the other. That held for two voices from the same engine, which disagree about time by a few percent. It would not have held for ANCIENT, which is a different engine and distributes time differently inside every phrase — measured, between 1.5 and 1.8 times the reference’s length, and never by the same amount twice.
So every overlay in OVERLORD is now locked to the reference’s delivery rather than to its punctuation. The engine measures where the syllables actually fall in the reference and in each overlay — the energy of the voice in five bands, and the slope where it rises — aligns the two by dynamic time warping, and then moves the overlay onto that path with a method that moves time and leaves pitch alone. Every onset in Jaimla, in ANCIENT and in LEADER lands on neural’s onset, and so does every onset in the other ten, within a grain of the voice; and because the lock does not touch pitch to move time, the octave Jaimla was measured onto stays an octave at every word rather than on average. The largest correction the lock had to make on the document is written into the recording’s manifest, in milliseconds, so “in time” is a figure and not a claim.
A recording that long is not swapped in until it is whole. The store renders a new OVERLORD beside the one that is playing and moves it into place only once its manifest is written, so what you hear under this name is always a finished recording: the previous one until the new one exists.
ANCIENT is a whisper with real tone underneath it, and a whisper does not begin the way a spoken line begins. Breath arrives before the first word and is still leaving after the last. Starting exactly on the first phoneme made it sound switched on rather than arrived, so the reading is now framed: the opening words and the closing words are rendered a second time in breath alone and laid before and after the voice, overlapping into each end. It is air and articulation only — it cannot form a word the voice is not already saying, which is exactly why it reads as the voice opening rather than as somebody else in the room.
ANCIENT was also speaking in the wrong accent, and had been for as long as it existed. Its voice file was rewritten onto a Gaelic-substrate English and said so on its second line; the entry that asked for it was still asking for Received Pronunciation — a broadcast accent standardised in the 1920s, and the youngest English in the cast, for a voice meant to predate it by four centuries. A synthesiser honours the accent it is handed and says nothing about the one the file wanted.
LEADER was built out of a lesson: it used to have a second voice mixed over it, and however that second voice was placed it was heard as a person — forming words where it overlapped the consonants and droning where it did not. So it was stripped back to one voice shaped by a broadcast EQ, and the note in its file says a second voice is not character, it is company.
That is true of a voice mixed dry. What makes it a person is the onset — the moment it starts, which the ear locates. Delay the same voice below the threshold where a delay is heard as a delay and its onsets arrive inside the body’s own; it stops being a person and becomes the voice’s own tail, and a tail is a timbre. So LEADER has a second face, at the same name and with the same body: Jaimla an octave above it at OVERLORD’s own ratio and entirely in echo, and under that, LEADER OF EARTH itself — the eight-accent build the name comes from — as tone, band limited below the consonants so it lends its mass without being able to say anything. Press LEADER twice; the label does not change, because the only honest indicator is the sound.
Savante is new, and she is derived from Jaimla rather than from the reference — the only voice here that is. Jaimla is the first voaice, so Savante’s body is hers and is not pitched: a ratio on a male voice does not make a woman, which is the correction Jaimla exists to embody. Everything else about Savante is the rest of this page applied at once — OVERLORD’s ratio, with the male half an octave below her and entirely in echo, so it is her depth rather than a second speaker; and ANCIENT’s frame, in a female breath, so she arrives and leaves rather than switching on and off. A savant does not raise her voice. Everything that would have been volume is depth instead.
Her depth is locked to her. The octave below used to be fitted to her at every comma and left as one uniform stretch between them, and two voices from the same engine still disagree about time inside a phrase by a few percent, which is enough to hear as someone a beat behind her. It now follows her delivery word by word, the same lock OVERLORD’s fourteen layers carry, so every onset of the resonance arrives inside her own and its pitch is not touched to get there.
Her name is said sav-ont — two syllables, the second carrying the weight — as in the old idiot savant, the one who cannot be taught and cannot be wrong about the thing she knows. She is the truth teller, and the master of mathematics. A synthesiser handed the written name read a third syllable into it, sav-ahn-tay, so the name is in the pronunciation table now, and every lane says it the same way: the rendered files, the reader in this browser, and the realtime engine. Savante is powered by sAGI — © 2026 PYTHAI.
LISTEN, beside the heading of this page, opens the voice panel and begins reading in the same gesture. It reads the document the browser is already showing, and it lights each word as it is said.
There are two ways it can do that, and the panel says which it is using. live means your own browser is speaking, on-device: nothing is downloaded and nothing leaves the page. file means this page was rendered ahead into an audio file, which is what you will hear if a recording exists in the voice you picked — it starts instantly, seeks, and can be downloaded. The player prefers the file, because someone who presses LISTEN twice wants the second press to be immediate.
When it is speaking live, the highlight follows the synthesiser's own word-boundary events rather than a timer; when it is playing a file, it follows marks measured at the moment the file was made. Either way the word that is lit is the word you are hearing, not an estimate of it.
To read something that is not this page, there is doc.player: paste any URL and hear it read, with the option to render it to a file you keep.
Open AUDIO DECK in the player and the controls appear a second time as instruments — a gain knob, a level meter, a speed control, a voice selector and a transport. It is the same machine with two faces: the knobs and the plain controls read and write the same values, so they cannot disagree, and the player works exactly as well with the deck never loaded.
The gain knob goes to 400%, and above 100% it is genuinely amplifying rather than just
un-attenuating. An <audio> element's own volume can only ever turn a sound
down — it clamps at 1.0 — so past that point a Web Audio gain stage carries the rest.
ANCIENT needs it: that voice is a whisper by design and is still the quietest thing in the cast
after loudness normalisation. The taper is squared, because loudness is not linear in the ear and
a linear knob wastes most of its travel where nothing much changes.
The instruments are DreamKnob
(dreamknob) — knobs, meters, segment displays and transports drawn as studio hardware,
which is the right shape for controls you adjust by feel rather than by number.
The JavaScript idiom here — plain modules, no framework on the page itself, one behaviour per file — follows javascriptit, and the interaction pieces (drag, drop, resize, the small reusable behaviours) follow mlodular. Both are the reference for what gets written by hand here and what gets reached for from a library.