diff --git a/docs/guides/voice-and-audio.mdx b/docs/guides/voice-and-audio.mdx
index 4687b19db4b..31ba8df0205 100644
--- a/docs/guides/voice-and-audio.mdx
+++ b/docs/guides/voice-and-audio.mdx
@@ -113,6 +113,11 @@ When narration drives the project, time visual changes to its real transcript
 rather than estimated scene lengths. When music drives the project, cut to its analyzed structure (see
 [music to video](/guides/music-to-video)).
 
+Music can start after narration begins. When the media-use skill creates a
+ducking lane for that clip, its opening gain matches the speech at that point
+in the composition. Starting during an attack or release continues that ramp
+from the correct level; completed speech does not create a new dip.
+
 ## Review the complete sound pass
 
 - The correct voice and source recordings are used.
diff --git a/packages/cli/src/media-use/lib/duck.mjs b/packages/cli/src/media-use/lib/duck.mjs
index dcf81774f39..8541f589687 100644
--- a/packages/cli/src/media-use/lib/duck.mjs
+++ b/packages/cli/src/media-use/lib/duck.mjs
@@ -63,11 +63,27 @@ export function duckLane(keyframes, { clipStart = 0, baseVolume = 1 } = {}) {
     if (t > points.at(-1).t) points.push({ t: round3(t), v });
   };
   for (const kf of keyframes) {
-    const t = Math.max(0, kf.time - start);
+    const t = Math.max(0, kf.time);
     push(t, points.at(-1).v);
     push(t + kf.duration, kf.volume);
   }
-  return { version: 1, lanes: [{ target: "volume", points }] };
+
+  const next = points.findIndex((point) => point.t > start);
+  const before = points[next - 1] ?? points.at(-1);
+  const after = points[next];
+  const initial =
+    next === 0
+      ? points[0].v
+      : after
+        ? before.v + ((after.v - before.v) * (start - before.t)) / (after.t - before.t)
+        : before.v;
+  const localPoints = [
+    { t: 0, v: initial },
+    ...points
+      .filter((point) => point.t > start)
+      .map((point) => ({ t: round3(point.t - start), v: point.v })),
+  ];
+  return { version: 1, lanes: [{ target: "volume", points: localPoints }] };
 }
 
 function mergeIntervals(intervals, mergeGap) {
diff --git a/packages/cli/src/media-use/lib/duck.test.mjs b/packages/cli/src/media-use/lib/duck.test.mjs
index 164c12bbdc5..dcb39a36235 100644
--- a/packages/cli/src/media-use/lib/duck.test.mjs
+++ b/packages/cli/src/media-use/lib/duck.test.mjs
@@ -169,6 +169,85 @@ test("duckLane turns composition-time keyframes into clip-local ramps", () => {
   });
 });
 
+for (const [name, clipStart, expected] of [
+  [
+    "before the attack",
+    0.5,
+    [
+      { t: 0, v: 0.6 },
+      { t: 0.5, v: 0.6 },
+      { t: 0.65, v: 0.15 },
+      { t: 2.5, v: 0.15 },
+      { t: 2.9, v: 0.6 },
+    ],
+  ],
+  [
+    "during the attack",
+    1.075,
+    [
+      { t: 0, v: 0.375 },
+      { t: 0.075, v: 0.15 },
+      { t: 1.925, v: 0.15 },
+      { t: 2.325, v: 0.6 },
+    ],
+  ],
+  [
+    "during speech",
+    2,
+    [
+      { t: 0, v: 0.15 },
+      { t: 1, v: 0.15 },
+      { t: 1.4, v: 0.6 },
+    ],
+  ],
+  [
+    "at the release",
+    3,
+    [
+      { t: 0, v: 0.15 },
+      { t: 0.4, v: 0.6 },
+    ],
+  ],
+  [
+    "during the release",
+    3.2,
+    [
+      { t: 0, v: 0.375 },
+      { t: 0.2, v: 0.6 },
+    ],
+  ],
+  ["after narration", 4, [{ t: 0, v: 0.6 }]],
+]) {
+  test(`duckLane preserves composition-time gain when music starts ${name}`, () => {
+    const keyframes = duckKeyframes([{ start: 1, end: 3 }], { baseVolume: 0.6 });
+    const points = duckLane(keyframes, { clipStart, baseVolume: 0.6 }).lanes[0].points;
+
+    assert.equal(points.length, expected.length);
+    for (let i = 0; i < points.length; i++) {
+      assert.equal(points[i].t, expected[i].t);
+      assert.ok(
+        Math.abs(points[i].v - expected[i].v) < 1e-9,
+        `gain at ${points[i].t}s: expected ${expected[i].v}, received ${points[i].v}`,
+      );
+    }
+  });
+}
+
+test("duckLane discards earlier narration and retains the next speech span", () => {
+  const keyframes = duckKeyframes([
+    { start: 1, end: 2 },
+    { start: 5, end: 6 },
+  ]);
+
+  assert.deepEqual(duckLane(keyframes, { clipStart: 3 }).lanes[0].points, [
+    { t: 0, v: 1 },
+    { t: 2, v: 1 },
+    { t: 2.15, v: 0.25 },
+    { t: 3, v: 0.25 },
+    { t: 3.4, v: 1 },
+  ]);
+});
+
 test("--json spans match --merge-gap semantics exactly", () => {
   const dir = mkdtempSync(join(tmpdir(), "media-use-duck-"));
   try {
diff --git a/skills-manifest.json b/skills-manifest.json
index e6442718b8a..6c88a545c50 100644
--- a/skills-manifest.json
+++ b/skills-manifest.json
@@ -18,7 +18,7 @@
       "files": 13
     },
     "media-use": {
-      "hash": "1082bd1c4e3f033b",
+      "hash": "6550641984b19843",
       "files": 116
     },
     "remotion-to-hyperframes": {
diff --git a/skills/media-use/scripts/lib/duck.mjs b/skills/media-use/scripts/lib/duck.mjs
index dcf81774f39..8541f589687 100644
--- a/skills/media-use/scripts/lib/duck.mjs
+++ b/skills/media-use/scripts/lib/duck.mjs
@@ -63,11 +63,27 @@ export function duckLane(keyframes, { clipStart = 0, baseVolume = 1 } = {}) {
     if (t > points.at(-1).t) points.push({ t: round3(t), v });
   };
   for (const kf of keyframes) {
-    const t = Math.max(0, kf.time - start);
+    const t = Math.max(0, kf.time);
     push(t, points.at(-1).v);
     push(t + kf.duration, kf.volume);
   }
-  return { version: 1, lanes: [{ target: "volume", points }] };
+
+  const next = points.findIndex((point) => point.t > start);
+  const before = points[next - 1] ?? points.at(-1);
+  const after = points[next];
+  const initial =
+    next === 0
+      ? points[0].v
+      : after
+        ? before.v + ((after.v - before.v) * (start - before.t)) / (after.t - before.t)
+        : before.v;
+  const localPoints = [
+    { t: 0, v: initial },
+    ...points
+      .filter((point) => point.t > start)
+      .map((point) => ({ t: round3(point.t - start), v: point.v })),
+  ];
+  return { version: 1, lanes: [{ target: "volume", points: localPoints }] };
 }
 
 function mergeIntervals(intervals, mergeGap) {