The speaker

Where it comes from. Valentina and Gulnura between 31-07 and 28-08-2026: thirty handovers, more than $120 of engine time, and one night that put a speaker on a live site with the wrong words in her mouth. Everything below was paid for once. The night, written up · the money, counted.

One cause under three faults. Jumping transitions, a rhythm that changes, a face that drifts in size: three seams, one reason. Every take is born as its own scene, capped at fifteen seconds, then stitched to its neighbours. On Gulnura's evening, 42 clips bought became 17 takes and 7 published parts: 520 seconds paid for 170 seconds on the page.

The route, settled. Reverse the order. Approve the voice first, cheaply, then birth each page section as ONE clip from the locked portrait plus that audio. No fifteen second cap, no seam inside a section, the same face throughout. Frank judged three versions of one real segment blind on 28-08-2026, a native take against two born on a rented machine from the same approved audio, and kept the machine route for new speakers: « the 3 are good », « good voice and good image ». Valentina stays on her native voice.

The number that comes with it. The machine hands the voice back about 14 decibels under where a page expects it, its own attenuation and not the recording: measured at −8.5 against −22.9. Frank heard it before anything was measured. Anything driving that engine levels the sound afterwards, which nacer.py does in one pass. The door for all of it is the speakers topic.

Rule 1. A band, never a ceiling. A take is accepted only between 1.85 and 2.10 words a second. The old rule rejected anything above 2.3, so the slow takes were redone and the brisk ones survived, and one visit mixed 1.75 with 2.17. The ear hears that seam.

Rule 2. When the rule changes, every take is reborn. Never only the failing ones. Sameness across a page beats saving a few dollars on the takes that scraped through.

Rule 3. Four layers, on the served page. What she says, transcribed from the clips the live site serves, equals the caption, equals the approved pack, and the pack matches the page sections. One command checks all four before and after every publish, so a colleague editing the page flags the affected parts within minutes. Her words passed every script check that night and still trailed the page, because no check tied the script to the screen. Working example: walk-served.py.

Rule 4. Freeze the text before recording. Recording starts only on a page whose text its owner has declared frozen. A later edit is not forbidden, it simply re-opens the affected segments and they rejoin the queue.

Rule 5. A ruler that invents a defect costs money. The transcriber is what decides whether a take worth a dollar is kept or bought again, and ours misreads: the same clip returned « Azfa », « AZFA » and « AFFA » on three passes, and heard relaciona as relación a. Calibrate every ruler against work already approved before it judges anything new, and give it the list of names it must not guess.

Rule 6. Read what renders, never the source. Any claim about what a visitor sees comes from the living page. A dead text field in the code produced a false alarm about the captions, and cost a day.

The free road, and when to leave it. Grok's own subscription session births takes at no cost per second: 32 takes for nothing on 21-08. The paid interface is $0.081 a second, and it is what $120 went to when the free road was down all night. Check the road before launching a batch. Recipe: grok-speaker.

The exception on voices, measured 24-08. The house rule says take a library voice rather than clone an invented person. It stops being true when one engine makes the face and the voice together. Same sentence, same image, a clone against five professional voices: Frank chose the engine's own voice three times out of three. It costs $0.65 a line against about two centimes, and it is paid because the ear asks for it, not because there is nothing else.

Still, she speaks but does not move. What Zalia does on the Comores leaflets and Grace in Lesotho: a photo and a recorded voice. Quick and almost free, only the voice costs.

Moving: she speaks and her face moves with the words. Valentina. Since 28-08-2026 the standard needs only a locked portrait plus her approved voice, born whole on a rented machine — no master clip: how, and what it costs. A speaker built before that date (a master clip, then a cheap repaint per part) keeps working the old way: that route, step 2.

Still speaker Speaks, does not move. Zalia · Grace · the AZFA circle.

Moving speaker Speaks and moves. Valentina, live.

How much of the person the page shows. Every size speaks.

Face To the bottom of the neck.

Half bust Head and shoulders only.

Bust To the chest, Valentina today. Live

Full body To the feet. The image is ready; the speaking version not yet.

The engines that make a face return a small picture, about 640 pixels. That size, not the mouth, is what makes a speaker look soft. Enlarging her four times and putting a film grain back fixes it - but it is slow, and it is the only slow step in the whole build. So it is a choice, and here is the same face under each one. Look first, read after.

Leave her as she is No waiting. What the AZFA page has shown since July.

The quick finish About 35 minutes of computer time per minute of speaker.

The slow finish About 13 hours per minute of speaker. Same enlargement, a heavier network inside.

What that means for a real page. Valentina speaks for two minutes and forty-five seconds on the AZFA page. That is under two hours of unattended computer time on the quick finish, and about forty hours on the slow one. Both are free, it is your machine, not a service.

Ask for it by name. Tell Claude Code “use the quick finish” or “use the slow finish”. If you say nothing, it uses the quick one and shows you the two side by side before spending a night on the other.

Measured 05-08-2026 on a Mac, on Valentina's own frames at 640 pixels: 1.5 seconds a frame against 29 to 35. Enlarging is the same ×4 in both; only the network differs. Three things that look like they would help and do not: cutting the frame into tiles (worse), half precision (no change), and a rented GPU (22 seconds a frame, and it cannot hand the file back).

How it works. The visitor asks, typed or spoken; an agent finds the answer in a small set of fact files we wrote, and speaks it back in the speaker's own voice. She answers only from those files, nothing invented, internal notes can never leak. This grounding is called RAG; our method is the voice-agent skill. A speaker who converses is the guide, and she has her own page: The guide (06-08-2026).

What is missing, for Valentina. A full voice test on the live page (a page opened from disk cannot use the microphone) · cutting her off mid-answer, never verified · the greeting is long in voice mode · the face does not answer, during a conversation the bust steps aside and only the voice replies, which is the real meaning of "not ready" · switching text to voice loses the dialogue · cost per minute unmeasured and the agent has no domain lock. Full detail: the diagnosis.

1. Portrait on green, one photo, head and shoulders, so she can be cut out.
2. Voice, reuse or clone; acronyms spelled as they must sound, locked in a dictionary.
3. Script in parts, the page's argument, 6 to 10 pieces of 15 to 30 seconds.
4. One master clip: once per person, the living eyes. Two routes now, Veo 3.1 Fast on our own key ($0.80) or Magic Hour (~1.20 EUR); never skip it, a free animate-a-still route gives flat, dead eyes. Both routes, drawn.
5. The factory, repaints only the mouth per recording, under a centime a part. Drawn, step by step.
6. The brief for a new speaker: the sheet you copy for each person, with what never changes (the flat green, the rhythm, one take per part). The template to copy.
7. Cut-out in pairs, one file for Chrome, one for Safari, always both.
8. Page wiring, idle breath, card, icons, the slow glide; then the verification walk.

Since 07-08-2026, the best-looking parts are born, not repainted. Frank rejected every processed version of a part and chose the raw clip three times running, so a part is now bought born: the engine speaks that exact line, at its own size, and the only step before the page is removing the green. About $1 a part, once. The factory stays for long narration and for people we filmed. The verdict, on the factory page · criteria 20 and 21.

Since 28-08-2026, step 4 is skippable. A locked portrait plus the approved voice is enough; the machine (nacer.py) births the whole section, no master clip, no repaint. Steps 1, 2, 3, 6, 7 and 8 still apply. Detail and cost: above.

All four were invisible on the desktop preview and all four hit Frank on his phone within minutes of the ABC note going live. Check them before saying a speaker is online.

1. The still picture must be a keyed PNG, never the portrait JPG. The page falls back to it whenever video cannot play, and a JPG still carries the green background: she appears inside a rectangle. Cut it out with the same recipe as the clips, and add a version mark to the file name or phones keep the old one.
2. The voice-only fallback must hand over to the video, not lock it out. The small sound file always wins the race on a phone; our player then refused the video that arrived a second later, so she stayed a still picture with a voice, forever. The video takes over at the same word. Test on a slow connection, never on the machine that made it.
3. Low Power Mode blocks even a silent video on iPhone, so she was simply absent and the card floated over nothing. If nothing has painted 1.5 seconds after the page loads, show the cut-out still; any video that starts hides it again.
4. A button's own label can swallow its tap. The stop control's word sits over the button once visible; on a phone the tap landed on the word. Let the visible label take clicks too, and make stop unconditional so no stale state can make it do nothing.

And "online" means the live address serves it. A push is not a deploy: rapid pushes queue behind each other and Frank read old text three times in a row. Wrap the publish so it only reports success after the live page proves the change.

Corrected 31-08-2026. This table used to price the old way: one master clip, then a mouth repainted on each part. Since the route was settled on 28-08 a section is born whole, so the money is counted per published minute, not per take. The door for all of it is the speakers topic.

Paid once, per person. A locked portrait, and her voice approved before any picture exists: cents a line at ElevenLabs, or nothing at all with Voicebox on the Mac. Getting the voice right first is the whole saving, because then a changed word is a re-render and not a purchase.

Then, per published minute, in one look:

Where the $33 comes from, corrected 31-08-2026. This table said $1 to $3 on 28-08, then $35 to $47 earlier today — both from a guessed hourly rate. The measurement, from the engine test's own record: two clips of 6.4 seconds, about 2.5 hours of A100 each; round 6 of that same test proved the cheaper A40 cannot even load the model. A published minute is 9.4 clips, so about 23 hours of A100, and runpod.io/pricing, checked 31-08-2026, prices that machine at $1.39 (PCIe) to $1.59 (SXM4) the hour: $33 to $37. The whole test, failed rounds included, cost $17 for 12.8 seconds of video.

So the machine route is the most expensive one we have, not the cheapest. It was chosen on 28-08 because Frank's eye picked it blind, and that reason stands on its own. But anything that plans a long narration on it should price it at $33 to $37 the minute, look again at HeyGen at $1 to $4, and remember that Grok through the subscription still costs nothing beyond the subscription.

One rule before any per-minute service: check what the open model on the rented computer does first, and say both prices. That is how the month that cost more than $120 ends.

This whole page is a skill, a small file of instructions their own assistant reads and follows: a person in the page who presents it in her own voice. They install it once; after that they simply say what they want, in their own words, and their assistant does it, on their own machine, with their own accounts, nothing of Frank's.

One command, in their Claude Code:

npx skills add gfrankgva/studio-skills

That installs this skill together with the studio's six others.

Or take the file itself: download the skill (one file, put it in their ~/.claude/skills/site-speaker/) · read it on GitHub · the whole pack.

Without the skill it still works, the method on this page still works by hand. The skill only means their assistant already knows it.