@ignix-social/editor 0.1.23 → 0.1.24

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ignix-social/editor",
-  "version": "0.1.23",
+  "version": "0.1.24",
   "description": "Ignix Editor: render HTML video compositions (reels, edits, clips) on your own computer. Based on HyperFrames.",
   "license": "Apache-2.0",
   "type": "module",

package/skills/ignix-editor/SKILL.md

@@ -13,7 +13,7 @@ Commands. `ignix-editor` means the copy the Ignix playbook installed: `~/.ignix-
 |---|---|
 | `ignix-editor doctor` | Node 22+ and the H.264 encoder; downloads its own ffmpeg and whisper the first time. Run it first. |
 | `ignix-editor init <name>` | Creates `~/ignix-videos/<name>` with a working skeleton, GSAP, local fonts and `ignix-frame.js`. |
-| `ignix-editor transcribe <file> --lang es [--script "…"] -o words.json` | Words with start/end seconds, on this machine. |
+| `ignix-editor transcribe <file> --lang es [--script "…"] -o words.json` | Words with start/end seconds, plus `pauses` ([start, end] of every silence ≥0.35 s, for cutting silences), on this machine. Over 2 min it prints whisper's progress; without a GPU it takes about as long as the audio. |
 | `ignix-editor faces <video> -o work/faces.json` | Where the face is over time (macOS and Windows; the first run downloads the face detector): a smoothed path `{t, cx, cy, size}` (0–1, top-left origin) to frame the person. |
 | `ignix-editor pick-audio [dir] --kind music\|sfx --query "inspiring" [--order trending]` | Music (or a specific sound effect) from Pixabay, chosen by the USER: opens Pixabay's search in their browser, waits for their download, copies it into `assets/` and notes its source in `work/audio-sources.json`. Tell the user first that you opened Pixabay for them to pick one. Never fetch pixabay.com yourself: it blocks robots. |
 | `ignix-editor giphy --search "thumbs up" [--kind gifs\|stickers\|clips]` · `ignix-editor giphy <link> [dir]` | GIFs, stickers and clips from GIPHY, chosen by the USER: the first opens GIPHY's search in their browser; they copy the link of the one they like (Share → Copy link) and the second brings it into `assets/`: a GIF as MP4, a clip as MP4 with its audio, a sticker as a transparent frame sheet (see section 2, GIPHY stickers). |

package/src/cli.mjs

@@ -12,7 +12,7 @@ import { HOME, PLATFORM, TOOLS, ensureTool, ensureWhisperModel, findTool } from
 import { linkMedia, prepare, projectHtml } from './prepare.mjs';
 import { validateHtml } from './validate.mjs';
 import { engineEnv, renderProject, runEngine } from './engine.mjs';
-import { alignToScript, transcribe } from './transcribe.mjs';
+import { alignToScript, fitToPauses, transcribe } from './transcribe.mjs';
 import { faces as detectFaces } from './analysis/detect.mjs';
 import { frames, probe } from './analysis/media.mjs';
 import { serve } from './analysis/serve.mjs';
@@ -215,16 +215,21 @@ const commands = {
     const ffprobe = await ensureTool('ffprobe', log);
     const whisper = await ensureTool('whisper', log);
     const model = await ensureWhisperModel(opts.model ?? 'medium', log);
-    let words = transcribe({ ffmpeg, whisper, model, input, language: opts.lang ?? 'es', threads: opts.threads });
-    if (opts.script) {
-      const dur = Number(spawnSync(ffprobe, ['-v', 'error', '-show_entries', 'format=duration', '-of', 'csv=p=0', input], { encoding: 'utf8' }).stdout.trim());
-      words = alignToScript(opts.script, words, dur);
-    }
-    const json = JSON.stringify({ language: opts.lang ?? 'es', words }, null, 1);
+    const dur = Number(spawnSync(ffprobe, ['-v', 'error', '-show_entries', 'format=duration', '-of', 'csv=p=0', input], { encoding: 'utf8' }).stdout.trim()) || 0;
+    const mmss = (x) => `${Math.floor(x / 60)}:${String(Math.round(x % 60)).padStart(2, '0')}`;
+    const t0 = Date.now();
+    // Aviso de duración ANTES de empezar: sin GPU, una charla larga tarda más o menos lo que dura.
+    if (dur > 120) log(`Transcribing ${mmss(dur)} of audio with the ${basename(model).replace(/^ggml-|\.bin$/g, '')} model. Without a GPU this can take about as long as the audio itself; whisper prints its progress below.`);
+    const r = transcribe({ ffmpeg, whisper, model, input, language: opts.lang ?? 'es', threads: opts.threads, progress: dur > 120 });
+    let words = r.words;
+    if (opts.script) words = alignToScript(opts.script, words, dur);
+    words = fitToPauses(words, r.pauses);
+    const json = JSON.stringify({ language: opts.lang ?? 'es', duration: +dur.toFixed(3), words, pauses: r.pauses }, null, 1);
     if (opts.output) {
       mkdirSync(dirname(resolve(opts.output)), { recursive: true });
       writeFileSync(resolve(opts.output), json);
     } else console.log(json);
+    log(`✓ ${words.length} words and ${r.pauses.length} pauses (≥0.35 s) in ${mmss((Date.now() - t0) / 1000)}${opts.output ? ` → ${opts.output}` : ''}`);
   },
 
   // Música o efectos de Pixabay elegidos por la persona (src/pick-audio.mjs): abre la
@@ -296,7 +301,7 @@ const commands = {
     if (r.sheet) {
       const sh = r.sheet;
       log(`✓ sticker → ${r.file}: ${sh.frames} frames in a ${sh.cols}×${sh.rows} transparent sheet, ${sh.frameWidth}×${sh.frameHeight} each, ${sh.fps} fps (${sh.durationSec} s loop). Animate it from t as the skill shows (GIPHY section).`);
-    } else log(`✓ ${r.kind} → ${r.file}${r.hasAudio ? ' (with its audio)' : ''}`);
+    } else log(`✓ ${r.kind} → ${r.file}${r.kind === 'clip' ? (r.hasAudio ? ' (with its audio)' : ' (this clip has no sound: treat it as a muted clip)') : ''}`);
     if (opts.json) console.log(JSON.stringify(r));
   },
 

package/src/giphy.mjs

@@ -67,7 +67,8 @@ export function giphySearchUrl(query, kind = 'gifs') {
 
 /** Qué fichero traer de media.giphy.com para cada tipo (y el de reserva). */
 export function giphyRenditions(kind) {
-  if (kind === 'clip') return [['giphy480p.mp4', 'mp4'], ['giphy.mp4', 'mp4']];
+  // El de 360p es el que lleva el audio; giphy480p.mp4 dejó de servirse (404, 2026-10-08).
+  if (kind === 'clip') return [['giphy360p.mp4', 'mp4'], ['giphy.mp4', 'mp4']];
   if (kind === 'sticker') return [['giphy.gif', 'gif']];
   return [['giphy.mp4', 'mp4'], ['giphy.gif', 'gif']];
 }
@@ -93,21 +94,21 @@ export function gifToSheet({ ffmpeg, ffprobe, gif, out }) {
 
 /**
  * El tipo cuando el enlace no lo dice (el del ARCHIVO, …/giphy.gif): con él un clip se traía
- * como GIF, sin audio. Solo los clips tienen giphy480p.mp4, y la página dice si es sticker.
+ * como GIF, sin audio. La página lo dice: los clips traen su bloque "video" y los stickers
+ * is_sticker.
  */
 export async function resolveGiphyKind(id, fromLink) {
   if (fromLink) return fromLink;
   if (!ID.test(id)) return 'gif';
-  const clip = await fetch(`https://media.giphy.com/media/${id}/giphy480p.mp4`, { method: 'HEAD', signal: AbortSignal.timeout(8000) })
-    .then((r) => r.ok)
-    .catch(() => false);
-  if (clip) return 'clip';
   const page = await fetch(`https://giphy.com/gifs/${id}`, { headers: { 'user-agent': 'Mozilla/5.0 (compatible; IgnixEditor)' }, signal: AbortSignal.timeout(8000) })
     .then((r) => (r.ok ? r.text() : ''))
     .catch(() => '');
   const html = page.replace(/\\"/g, '"');
   const at = html.indexOf(`"id":"${id}"`);
-  return at >= 0 && /"is_sticker":true/.test(html.slice(at, at + 8000)) ? 'sticker' : 'gif';
+  if (at < 0) return 'gif';
+  const next = html.slice(at + 10).search(/"id":"[A-Za-z0-9]{8,40}"/);
+  const own = html.slice(at, next < 0 ? at + 8000 : at + 10 + next);
+  return /"video":\{/.test(own) ? 'clip' : /"is_sticker":true/.test(own) ? 'sticker' : 'gif';
 }
 
 /** Trae el elegido a `<project>/assets/` y apunta su origen en `work/media-sources.json`. */
@@ -143,7 +144,12 @@ export async function importGiphy({ link, project, kind: forced, ffmpeg, ffprobe
       /* fichero roto: se empieza de nuevo */
     }
     writeFileSync(log, JSON.stringify([...prev, { source: 'giphy', kind, id: parsed.id, link: String(link), file: dest, at: new Date().toISOString() }], null, 2));
-    return { id: parsed.id, kind, file: dest, hasAudio: kind === 'clip' && file.includes('480p'), ...(sheet ? { sheet } : {}) };
+    // Audio de verdad, no supuesto: hay clips mudos («Take My Money»).
+    const hasAudio =
+      kind === 'clip' && ffprobe
+        ? spawnSync(ffprobe, ['-v', 'error', '-select_streams', 'a', '-show_entries', 'stream=index', '-of', 'csv=p=0', dest], { encoding: 'utf8' }).stdout.trim() !== ''
+        : false;
+    return { id: parsed.id, kind, file: dest, hasAudio, ...(sheet ? { sheet } : {}) };
   }
   throw new Error(`Could not download that ${kind} from GIPHY${lastErr ? `: ${lastErr.message}` : ''}.`);
 }

package/src/transcribe.mjs

@@ -46,19 +46,60 @@ function whisperTimeoutMs(wavPath) {
   } catch {
     /* sin tamaño: el mínimo */
   }
-  return Math.max(10 * 60_000, seconds * 6 * 1000);
+  // ENTERO: spawnSync rechaza un timeout con decimales («The value of "timeout" is out of
+  // range… Received 13158197.625») y eso rompía cualquier audio de más de 100 s (2026-10-08).
+  return Math.ceil(Math.max(10 * 60_000, seconds * 6 * 1000));
+}
+
+/**
+ * Pausas del audio (ffmpeg silencedetect): [[inicio, fin], …] en segundos. Sirven para el
+ * paso de silencios y para que ninguna palabra se estire sobre una pausa: whisper con -ml 1
+ * hace que cada palabra acabe donde empieza la siguiente («claro.» de 0,76 s, «1.600,» de 1,6 s).
+ */
+export function detectPauses(ffmpeg, input, { noiseDb = -35, minSec = 0.35 } = {}) {
+  const r = spawnSync(ffmpeg, ['-nostdin', '-hide_banner', '-i', input, '-vn', '-af', `silencedetect=noise=${noiseDb}dB:d=${minSec}`, '-f', 'null', '-'], { encoding: 'utf8', maxBuffer: 64 * 1024 * 1024 });
+  const out = [];
+  let start = null;
+  for (const line of (r.stderr || '').split('\n')) {
+    const a = line.match(/silence_start: (-?[\d.]+)/);
+    if (a) start = Math.max(0, Number(a[1]));
+    const b = line.match(/silence_end: ([\d.]+)/);
+    if (b && start !== null) {
+      out.push([+start.toFixed(3), +Number(b[1]).toFixed(3)]);
+      start = null;
+    }
+  }
+  return out;
+}
+
+/** Recorta cada palabra a la voz: ni empieza ni acaba dentro de una pausa (si le queda voz). */
+export function fitToPauses(words, pauses) {
+  if (!pauses?.length) return words;
+  return words.map((w) => {
+    let { s, e } = w;
+    for (const [ps, pe] of pauses) {
+      if (ps >= e || pe <= s) continue;
+      if (ps > s && ps - s >= 0.08) e = Math.min(e, ps); // se estiraba sobre la pausa
+      else if (pe < e && e - pe >= 0.08) s = Math.max(s, pe); // empezaba dentro de la pausa
+    }
+    return { ...w, s: +s.toFixed(3), e: +e.toFixed(3) };
+  });
 }
 
 function whisperFailure(w, timeoutMs) {
   if (w.error?.code === 'ETIMEDOUT') {
     return new Error(`whisper did not finish in ${Math.round(timeoutMs / 60_000)} min and was stopped. Close other heavy programs and run the command again, or use a smaller model (--model small).`);
   }
-  return new Error(`whisper failed: ${(w.stderr || '').slice(-400)}`);
+  return new Error(`whisper failed (exit ${w.status ?? w.error?.code ?? '?'}): ${(w.stderr || '').slice(-400)}`);
 }
 
 const norm = (s) => s.toLowerCase().normalize('NFD').replace(/[̀-ͯ]/g, '').replace(/[^\p{L}\p{N}]/gu, '');
 
-export function transcribe({ ffmpeg, whisper, model, input, language = 'es', threads }) {
+/**
+ * Palabras y pausas. Con `progress`, whisper escribe su avance («progress = 40%») en la
+ * consola: en una transcripción larga, sin eso nadie distingue «va lento» de «se ha colgado».
+ */
+export function transcribe({ ffmpeg, whisper, model, input, language = 'es', threads, progress = false }) {
   const ws = whisperWorkspace(model);
   const tmp = ws.tmp;
   try {
@@ -67,8 +108,10 @@ export function transcribe({ ffmpeg, whisper, model, input, language = 'es', thr
     if (a.status !== 0) throw new Error(`Could not read the audio of ${input}: ${a.stderr.slice(-300)}`);
     const out = join(tmp, 'out');
     const args = ['-m', ws.modelArg, '-l', language, '-ml', '1', '-sow', '-oj', '-of', ws.arg(out), '-t', String(whisperThreads(threads)), '-f', ws.arg(wav)];
+    if (progress) args.unshift('-np', '-pp');
     const timeout = whisperTimeoutMs(wav);
-    const run = (extra) => spawnSync(whisper, [...extra, ...args], { encoding: 'utf8', maxBuffer: 64 * 1024 * 1024, cwd: ws.cwd, timeout });
+    const stdio = progress ? ['ignore', 'pipe', 'inherit'] : undefined;
+    const run = (extra) => spawnSync(whisper, [...extra, ...args], { encoding: 'utf8', maxBuffer: 64 * 1024 * 1024, cwd: ws.cwd, timeout, stdio });
     // En Mac Intel la GPU (Metal) se cuelga con whisper: CPU directamente. En el resto, GPU y,
     // si falla, otra vez por CPU.
     const intelMac = process.platform === 'darwin' && process.arch === 'x64';
@@ -82,7 +125,7 @@ export function transcribe({ ffmpeg, whisper, model, input, language = 'es', thr
       if (!t || /^\[.*\]$/.test(t)) continue; // [MÚSICA], [MUSIC]…
       words.push({ t, s: seg.offsets.from / 1000, e: seg.offsets.to / 1000 });
     }
-    return words;
+    return { words, pauses: detectPauses(ffmpeg, wav) };
   } finally {
     rmSync(tmp, { recursive: true, force: true });
   }