# Skill: full/speech-dub — "נאום לריל בעברית — הדקה הכי חזקה, מדובבת עם רגש" (Complete video) A long talk becomes a one-minute reel in Hebrew. Gemini watches the whole speech where it is on YouTube — nothing is downloaded — and picks its strongest continuous minute; the sentences are transcribed exactly with how each one is said, translated into spoken Hebrew written so a voice cannot misread it, and read line by line by ElevenLabs v3 with acting directions: calm in the story, a shout at the peak. Every word is checked by a transcript and the weak lines are re-recorded. Epic music under the voice, licensed stock footage over the whole video, kinetic words, black cards on the two strongest lines and the speaker credited on screen — the speech itself never appears. The example: Arnold Schwarzenegger's 45-minute "Rules of Success" → 81 seconds in Hebrew, in the owner's own cloned voice. For talks you may quote: your own, public-domain or Creative Commons speeches, or a short credited quote. **Includes:** a long talk on YouTube, watched where it is — nothing downloaded · its strongest continuous minute, chosen from the whole speech · the exact sentences, with how each one is said · a spoken translation, written so a voice cannot misread it · every line voiced with acting directions — calm in the story, a shout at the peak · every word checked by a transcript, weak lines re-recorded · epic music under the voice · licensed stock footage over the whole video · kinetic words, black cards on the two strongest lines, the speaker credited **Built on:** `POST /v1/ai/text` watching the YouTube link (no download), ElevenLabs v3 lines with direction tags joined in `voiceover` on the black canvas, the kinetic planner on the track's words and `POST /v1/stock/broll-plan` (faceless) for the picture, a `banner` title and a two-line credit **Price:** 290–430 credits per finished minute · 30 s ≈ 180–270 credits **Uses:** `/v1/ai/text` · `/v1/ai/voice` · `/v1/studio/edit` · `/v1/speech/transcripts` · `/v1/ai/music` · `/v1/studio/kinetic` · `/v1/stock/broll-plan` Base URL: `https://apihub.co.il` · Auth: `Authorization: Bearer ` · Docs: `/docs` Send the same `X-Tag: ` header on every call of one video; `GET /v1/usage?tag=` then returns exactly what it cost. Times in request bodies are **seconds on the edited (output) timeline**. Unknown fields are rejected with the name of the right field — read the error instead of guessing. **Files are kept 24 hours** (renders, uploads, generated clips/images/music) — download the finished video. **Models.** Gemini 3.1 Pro (`quality: "best"`) watches the whole talk at 0.2 fps, then the candidates at 1 fps, and listens to the delivery; ElevenLabs v3 reads the lines and makes the music; AssemblyAI checks every word; Gemini Flash-Lite picks and checks the stock (Pexels, Pixabay, CC0). The watch is priced by the talk's length, everything else by the reel's. --- ## 1. Watch the whole speech and pick its strongest minute ```jsonc POST /v1/ai/text { "quality": "best", "json": true, "max_tokens": 24000, "prompt": "Watch this whole speech. Find the single most motivating CONTINUOUS stretch of about 55–75 seconds that works on its own as a reel: a strong opening line, rising energy, a peak, a punchy last line, no setup that needs earlier context, sentence boundaries at both ends. Also give 2 alternatives. JSON: { best: { start, end, why, first_line, last_line }, alternatives: […], speaker_notes }", "media": [ { "url": "https://www.youtube.com/watch?v=", "type": "video", "start": 0, "fps": 0.2 } ] } // a 45-minute speech ≈ 100,000 tokens ≈ 96 credits, ≈ 2 min · "start": 0 matters — a link with &t= starts there ``` **Nothing is downloaded.** Gemini watches the link where it is; the reel shows licensed stock footage, never the speech, and names the speaker on screen. The words still belong to whoever owns the talk: your own talks, public-domain or Creative Commons speeches, or a short credited quote. **Low fps on purpose.** A speech is heard, not watched — 0.2 frames a second is enough to choose, and five times cheaper. ## 2. The exact sentences of the candidates ```jsonc POST /v1/ai/text { "quality": "best", "json": true, "temperature": 0.1, "prompt": "Transcribe EVERY spoken sentence verbatim with its start and end in seconds of the FULL video (the clip starts at ). Mark laughter and applause as their own items. For each sentence add \"delivery\" (calm, joking, intense, shouting…). Rate the stretch 1–10 as a 60-second reel and say which sentence it should start and end on.", "media": [ { "url": "", "type": "video", "start": , "end": , "fps": 1 } ] } // ≈ 30–40 credits a candidate · run the best and the strongest alternative side by side, then choose ``` The model's first pick is often the most *viral* minute, not the most *motivating* one. In the example it chose "sleep faster" (a joke and the arithmetic of a day); the transcripts showed "don't listen to the naysayers" climbing from a story to a shout. Read both and choose. ## 3. The translation — written for a voice One line per sentence, each with `at` = where the speaker starts it. Your own pass, before any voice reads it: - **Spoken, not translated.** "I'm gonna be the one" is «אני אהיה הראשון», not «אני אהיה האחד הזה»; "I am a strong believer what Mandela said" is «אני חי לפי משפט אחד… המשפט של נלסון מנדלה». - **No em-dashes.** ElevenLabs v3 reads «—» as a dead stop; a comma keeps the flow. «אבל אני — הולך — לעשות את זה» came out robotic; «אבל אני הולך לעשות את זה!» is the peak. - **Words the voice stresses wrong** go: «סביבי», «אתלטיקה» → «בריצה», «אוכיח» → «וכולם עוד יראו». Quotes in ״…״. - **Applause** splits a sentence the speaker repeats ("Because I am a strong believer…" [applause] "I am a strong believer…"): keep both halves, the second picks up the thought. ## 4. One take per line, with direction ```jsonc POST /v1/ai/voice { "text": "[intense] אל תקשיבו לאלה שאומרים לכם ״אי אפשר״.", "voice_id": "", "stability": 0, "language_code": "he" } // ≈ 2–4 credits a line · → output.url, output.duration_seconds ``` - **The director does this for you on plain Hebrew** (`prepare`, on by default — tags, niqqud on every Hebrew word, stability from the words; say the delivery in `mood`). Write the tags yourself only when you want a specific one on a specific line; a text with tags or niqqud is read exactly as sent. - **Direction tags** in English at the start: `[storytelling]` `[serious]` `[sarcastic]` `[intense]` `[excited]` `[passionate]` `[shouting]` `[firm]`. `stability`: 0 for the emotional lines (more acting), 0.5 for the narration. - **The voice.** A voice from `GET /v1/ai/voices`, or your own: `POST /v1/ai/voices/clone` { "name", "audio_urls", "consent" } → a `voice_id` only your account can use (the example is the owner's own clone, made from ≈ 2 minutes of his WhatsApp videos; the default model, eleven_v4, is the newest and most emotive — never name an older one). A HeyGen clone (`POST /v1/voices` + `/v1/voices/speech`) is the speaker's timbre but reads flat — no direction tags. ## 5. The track Back to back on the black canvas, the speaker's own pauses tightened: ≈ 0.35 s between lines, 0.6–0.9 s before a peak. ```jsonc POST /v1/studio/edit { "video_url": "https://media.gpt.org.il/samples/canvas-9x16-black-15min.mp4", "trim": { "end": }, "audio": { "mute": true, "normalize": true }, "voiceover": [ { "url": "", "start": 0.6 }, { "url": "", "start": 4.65 }, … ], "output": { "format": "mp3" } } // ≈ 0.2 credits a second (81 s: 16) ``` ## 6. Check every word, then the delivery ```jsonc POST /v1/speech/transcripts { "audio_url": "", "language_code": "he", "speech_models": ["universal-2"] } // ≈ 2 credits ``` **The words:** read the transcript against your script, line by line — a misread word shows up as a different word. **The delivery:** `POST /v1/ai/text` (`quality: "best"`) with the track as `media` ("a native dubbing director: per line the emotion 1–10, robotic pauses, wording that sounds translated, the three weakest lines and how to direct them"). Never hand it a list of words to look for — it reports those words as misread even when they are no longer in the script; it also judges acting worse than an ear does. Re-record the weak lines (step 4, a new idempotency key), rebuild the track. ## 7. The music ```jsonc POST /v1/ai/music { "prompt": "Epic cinematic motivational underscore, 85 BPM, a simple hopeful piano motif over warm strings, soft taiko pulses entering after 30 seconds, building to a triumphant climb in the last 20 seconds, ends on one big resolved hit, instrumental, no vocals, stays under a speaking voice.", "duration": , "instrumental": true } // ≈ 0.95 credits a second ``` ## 8. The words and the stock ```jsonc POST /v1/studio/kinetic { "video_url": "https://media.gpt.org.il/samples/canvas-9x16-black-15min.mp4", "words": , "cut": false, "source": { "duration": , "fps": 30 }, "emphasis": [ … ], "scenes": [ { "word": }, { "word": } ], "music": { "url": "", "volume": 0.16 } } // 1 credit · the voice ends ≈ 16–19 dB over the bed POST /v1/stock/broll-plan { "style": "faceless", "words": , "duration": , "edit": , "voice": { "url": "", "seconds": }, "kind": "any", "verify": true, "moments": [ { "start": 0, "query": "man lifting heavy barbell", "short": "barbell" }, { "start": 23.7, "query": "skier downhill snow mountains", "short": "skiing" }, … ] } // ≈ 12 credits ``` One moment per sentence or two, an English search that shows what the line *means* (the parents saying no → an empty gym at night; "I'll be the first" → a man running up city stairs). No night skies or dark rooms — a nearly black shot reads as a hole in the video. Two black cards (`scenes`) on the strongest lines: the quote and the peak. ## 9. The title, the credit, one render The kinetic plan uses most of the 80 `texts` — the title goes in as a graphic, the credit as one text with two lines: ```jsonc "graphics": [ …, { "kind": "banner", "text": "", "sub": "״״ בעברית", "y": 21, "background": "#0B0B14", "enter": "pop", "during": { "from": "w0001", "to": "w0010" } } ], "texts": [ …, { "text": "נערך עם *apihub.co.il*\n · ", "preset": "pill", "effect": "box", "style": { "y": 92.5, "size": 3.4, "weight": 800, "background": "#0B0B14E6", "accent_color": "#D0F808" }, "during": { … the whole video } } ], "output": { "aspect_ratio": "9:16", "resolution": "720", "quality": "high" } POST /v1/studio/edit // ≈ 1 credit a second at 720p · x-wait-ms: 110000, then poll ``` **720p, not 1080p.** The stock is 720p, and 30+ shots in one 1080p render ran out of memory on the render server in the example. When you publish, credit the footage: "Videos provided by Pexels". ## What the example cost Arnold Schwarzenegger, "Rules of Success" (45 minutes) → 81 seconds in Hebrew (dev, 1.10.2026): | Step | Credits | |---|---| | Watch the whole speech, pick the minute | 96 | | The exact sentences of two candidates | 65 | | 17 lines in the cloned voice + 10 re-recorded | 72 | | The checks: two listening passes and the transcript | 56 | | The music, 82 s | 77 | | 34 stock shots (Pexels), chosen and checked | 12 | | The words: kinetic plan, transcripts, uploads | 7 | | Render, 720p | 77 | | **Total** | **462** | The example's track was joined on our side; through step 5 it adds ≈ 16. Without the second candidate, the listening passes and the re-recorded lines it is ≈ 370. The watch follows the SOURCE's length (≈ 2 credits a minute of it), the rest the reel's. Send `x-wait-ms: 110000` to wait for the render in the same call; if the response is still `processing`, poll `GET /v1/executions/{id}` until `status` is `completed` and read `output.url`. ## Check the bill `GET /v1/usage?tag=` — per call, per operation and the total, in credits. Quote that to the customer, not the estimate above. ## Rules that keep captions readable on every platform - **Safe zone.** Keep text inside the middle 80% of the frame: `side_margin` ≥ 6 handles the sides; `margin` ≥ 12 keeps the bottom clear of the platform UI (username, buttons, progress bar). - **A line never leaves the frame** — a line too wide for the side margins is set smaller automatically. Do not shrink the size yourself, wrap instead (`max_chars_per_line`). - **Hebrew has no capitals** — keep `uppercase: false`; it only breaks Latin words mixed into a sentence. - **Fonts.** `GET /v1/studio/fonts` lists the bundled families and which scripts they draw. A font that cannot draw the caption's language is swapped and reported in `warnings`. - The customer's transcript wins: to fix a word, `GET /v1/speech/transcripts/{id}/captions`, edit the text, and send the list back as `captions.captions` — timing is kept when the word count is unchanged.