diff --git a/ui/src/lib/sceneRecipe.ts b/ui/src/lib/sceneRecipe.ts
index 8c184fd2..123ed8f3 100644
--- a/ui/src/lib/sceneRecipe.ts
+++ b/ui/src/lib/sceneRecipe.ts
@@ -1290,6 +1290,25 @@ export function recipeAssetDuration(recipe: SceneRecipe, assetId: string): numbe
   return Math.max(0.5, ...(durations.length ? durations : [recipe.scene.duration || 5]))
 }
 
+/** TTS/SFX/music generation length for one recipe track.
+ *
+ * `scene.duration` is only a per-shot fallback (often the first hold). Qwen3
+ * and the other audio engines treat `duration_seconds` as a hard cap, so a
+ * later 9s spoken shot must not inherit an 6–8s scene default or the line is
+ * truncated while mouth keyframes keep moving.
+ */
+export function recipeAudioDuration(recipe: SceneRecipe, trackId: string): number {
+  const shots = listRecipeShots(recipe)
+  const fromShots = shots
+    .filter(shot => shot.audioTrackIds === undefined || shot.audioTrackIds.includes(trackId))
+    .map(shot => shot.duration || recipe.scene.duration || 5)
+  const fromBeats = (recipe.dialogueBeats ?? [])
+    .filter(beat => beat.audioTrackId === trackId)
+    .map(beat => beat.end)
+  const fallback = recipe.scene.duration || 5
+  return Math.max(0.5, ...(fromShots.length ? fromShots : [fallback]), ...fromBeats)
+}
+
 function scopeRecipeToShot(recipe: SceneRecipe, shot: SceneRecipeShot): SceneRecipe {
   const audioIds = shot.audioTrackIds === undefined ? undefined : new Set(shot.audioTrackIds)
   const beatIds = shot.dialogueBeatIds === undefined ? undefined : new Set(shot.dialogueBeatIds)
diff --git a/ui/src/lib/sceneRecipeAssets.ts b/ui/src/lib/sceneRecipeAssets.ts
index 4e0e2d90..2c2b70b5 100644
--- a/ui/src/lib/sceneRecipeAssets.ts
+++ b/ui/src/lib/sceneRecipeAssets.ts
@@ -1,7 +1,7 @@
 import * as api from '../api/client'
 import { generateImageAsset } from './imageGeneration'
 import type { SceneRecipe, SceneRecipeAsset, SceneRecipeAudio } from './sceneRecipe'
-import { aspectRatioForScene, h3FramesForDuration, h3ResolutionForScene, recipeAssetDuration } from './sceneRecipe'
+import { aspectRatioForScene, h3FramesForDuration, h3ResolutionForScene, recipeAssetDuration, recipeAudioDuration } from './sceneRecipe'
 
 const wait = (ms: number) => new Promise(resolve => window.setTimeout(resolve, ms))
 
@@ -210,7 +210,7 @@ async function resolveAudio(
   if (track.source) return track.source
   await waitForGpuIdle(onStatus, signal)
   onStatus?.(`Generating ${track.kind} “${track.id}”…`)
-  const started = await api.submitGeneration(recipeAudioGenerationParams(track, recipe.scene.duration || 5, workspace))
+  const started = await api.submitGeneration(recipeAudioGenerationParams(track, recipeAudioDuration(recipe, track.id), workspace))
   const status = await pollUntil(
     `${track.kind === 'speech' ? 'Voice' : track.kind === 'sfx' ? 'SFX' : 'Music'} “${track.id}”`,
     () => api.fetchJobStatus(started.job_id),
diff --git a/ui/tests/characterKitEpisode.test.mjs b/ui/tests/characterKitEpisode.test.mjs
index 66f22405..e6620ff9 100644
--- a/ui/tests/characterKitEpisode.test.mjs
+++ b/ui/tests/characterKitEpisode.test.mjs
@@ -2,7 +2,7 @@ import assert from 'node:assert/strict'
 import test from 'node:test'
 import { characterKitRecipeInventory, createCharacterKit, mountCharacterKitLayers } from '../src/lib/characterKit.ts'
 import { composeCharacterKitLook, lockFaceRigMouthPlacement } from '../src/lib/characterKitFaceRig.ts'
-import { compileRecipeShot, listRecipeShots, parseSceneRecipe } from '../src/lib/sceneRecipe.ts'
+import { compileRecipeShot, listRecipeShots, parseSceneRecipe, recipeAudioDuration } from '../src/lib/sceneRecipe.ts'
 import { evaluateSceneLayer } from '../src/lib/sceneTimeline.ts'
 
 const asset = (id, source, reviewState = 'approved', kind = 'overlay') => ({
@@ -129,6 +129,9 @@ test('CharacterKit episode recipe mounts only approved pieces and isolates audio
   const pointing = compileRecipeShot(recipe, shots[3], {}, filename => filename)
   assert.ok(pointing.layers.some(layer => layer.id === 'kit-luma-pose-pointing'))
   assert.deepEqual(pointing.audioTracks.map(track => track.id), ['voice-luma-2'])
+  assert.equal(recipe.scene.duration, 8)
+  assert.equal(shots[3].duration, 9)
+  assert.equal(recipeAudioDuration(recipe, 'voice-luma-2'), 9)
 })
 
 test('mini South Park-style cutout dialogue locks mouths and scopes speech per shot', () => {
diff --git a/ui/tests/sceneRecipe.test.mjs b/ui/tests/sceneRecipe.test.mjs
index aa4c90cf..51019731 100644
--- a/ui/tests/sceneRecipe.test.mjs
+++ b/ui/tests/sceneRecipe.test.mjs
@@ -14,6 +14,7 @@ import {
   parseSceneRecipe,
   parseSceneRecipeText,
   recipeAssetDuration,
+  recipeAudioDuration,
 } from '../src/lib/sceneRecipe.ts'
 
 test('example saucer recipe compiles to a 4-layer scene with space-cruise motion', () => {
@@ -395,6 +396,46 @@ test('H3 plates use supported model canvases and enough temporal-grid frames', (
   assert.equal(recipeAssetDuration(recipe, 'clouds'), 9)
 })
 
+test('recipe speech uses the spoken shot length, not the shorter scene default', () => {
+  const layers = [
+    { id: 'camera', type: 'camera', cameraPreset: 'camera-locked' },
+    { id: 'hero', type: 'image', asset: 'hero-art' },
+    { id: 'mouth-wide', name: 'Mouth wide', type: 'overlay', asset: 'mouth-art', faceBinding: { poseLayerId: 'hero', role: 'mouth', state: 'wide' } },
+  ]
+  const recipe = parseSceneRecipe({
+    version: 1,
+    name: 'episode-hold-then-line',
+    assets: [
+      { id: 'hero-art', kind: 'image', source: 'hero.png' },
+      { id: 'mouth-art', kind: 'image', source: 'mouth.png' },
+    ],
+    audio: [
+      { id: 'voice-snowman', kind: 'speech', prompt: 'El timbre de verdad está detrás del muñeco de nieve.' },
+      { id: 'voice-bell', kind: 'speech', prompt: 'La campana del patio está congelada.' },
+    ],
+    dialogueBeats: [
+      { id: 'beat-snowman', text: 'El timbre de verdad está detrás del muñeco de nieve.', start: 0.3, end: 7.6, mouthLayerIds: ['mouth-wide'], audioTrackId: 'voice-snowman' },
+      { id: 'beat-bell', text: 'La campana del patio está congelada.', start: 0.4, end: 6.8, mouthLayerIds: ['mouth-wide'], audioTrackId: 'voice-bell' },
+    ],
+    shots: [
+      { name: 'hold', duration: 6, audioTrackIds: [], dialogueBeatIds: [], layers },
+      { name: 'bell', duration: 8, audioTrackIds: ['voice-bell'], dialogueBeatIds: ['beat-bell'], layers },
+      { name: 'snowman', duration: 9, audioTrackIds: ['voice-snowman'], dialogueBeatIds: ['beat-snowman'], layers },
+    ],
+    scene: { width: 1280, height: 720, fps: 30, duration: 8, layers },
+  })
+  assert.equal(recipe.scene.duration, 8)
+  assert.equal(recipeAudioDuration(recipe, 'voice-snowman'), 9)
+  assert.equal(recipeAudioDuration(recipe, 'voice-bell'), 8)
+
+  const omittedSceneDuration = parseSceneRecipe({
+    ...recipe,
+    scene: { width: 1280, height: 720, fps: 30, layers },
+  })
+  assert.equal(omittedSceneDuration.scene.duration, 6)
+  assert.equal(recipeAudioDuration(omittedSceneDuration, 'voice-snowman'), 9)
+})
+
 const gradedRecipe = grade => parseSceneRecipe({
   ...EXAMPLE_SAUCER_CRUISE_RECIPE,
   scene: { ...EXAMPLE_SAUCER_CRUISE_RECIPE.scene, ...grade },
diff --git a/ui/tests/sceneRecipeAssets.test.mjs b/ui/tests/sceneRecipeAssets.test.mjs
index 2a963808..e8bebdaa 100644
--- a/ui/tests/sceneRecipeAssets.test.mjs
+++ b/ui/tests/sceneRecipeAssets.test.mjs
@@ -1,7 +1,7 @@
 import assert from 'node:assert/strict'
 import test from 'node:test'
 import { recipeAudioGenerationParams } from '../src/lib/sceneRecipeAssets.ts'
-import { withResolvedSources } from '../src/lib/sceneRecipe.ts'
+import { parseSceneRecipe, recipeAudioDuration, withResolvedSources } from '../src/lib/sceneRecipe.ts'
 
 test('recipe speech uses the proven HocusPocus voice-generation contract', () => {
   const params = recipeAudioGenerationParams({
@@ -53,3 +53,35 @@ test('recipe audio generation refuses an unresolved silent prompt', () => {
     /needs a prompt or an existing source/,
   )
 })
+
+test('auto-generated episode speech is capped at the spoken shot, not scene.duration', () => {
+  const layers = [
+    { id: 'camera', type: 'camera', cameraPreset: 'camera-locked' },
+    { id: 'hero', type: 'image', asset: 'hero-art' },
+    { id: 'mouth-wide', name: 'Mouth wide', type: 'overlay', asset: 'mouth-art', faceBinding: { poseLayerId: 'hero', role: 'mouth', state: 'wide' } },
+  ]
+  const recipe = parseSceneRecipe({
+    version: 1,
+    name: 'episode-hold-then-line',
+    assets: [
+      { id: 'hero-art', kind: 'image', source: 'hero.png' },
+      { id: 'mouth-art', kind: 'image', source: 'mouth.png' },
+    ],
+    audio: [
+      { id: 'voice-snowman', kind: 'speech', prompt: 'El timbre de verdad está detrás del muñeco de nieve.' },
+    ],
+    dialogueBeats: [
+      { id: 'beat-snowman', text: 'El timbre de verdad está detrás del muñeco de nieve.', start: 0.3, end: 7.6, mouthLayerIds: ['mouth-wide'], audioTrackId: 'voice-snowman' },
+    ],
+    shots: [
+      { name: 'hold', duration: 6, audioTrackIds: [], dialogueBeatIds: [], layers },
+      { name: 'snowman', duration: 9, audioTrackIds: ['voice-snowman'], dialogueBeatIds: ['beat-snowman'], layers },
+    ],
+    scene: { width: 1280, height: 720, fps: 30, duration: 8, layers },
+  })
+  const seconds = recipeAudioDuration(recipe, 'voice-snowman')
+  assert.equal(seconds, 9)
+  const params = recipeAudioGenerationParams(recipe.audio[0], seconds, 'episode-room')
+  assert.equal(params.duration_seconds, 9)
+  assert.notEqual(params.duration_seconds, recipe.scene.duration)
+})