diff --git a/.changeset/merge-frames-linear.md b/.changeset/merge-frames-linear.md
new file mode 100644
index 000000000..07490081b
--- /dev/null
+++ b/.changeset/merge-frames-linear.md
@@ -0,0 +1,5 @@
+---
+'@livekit/agents': patch
+---
+
+Make `mergeFrames` linear instead of quadratic. It previously rebuilt the accumulated buffer with a boxed spread on every frame (`new Int16Array([...data, ...frame.data])`), costing O(total_samples x frame_count) copies. It now sums lengths in one pass and copies each frame once into a preallocated `Int16Array`. Merging a 30s utterance of 10ms frames drops from ~60s of blocked event loop to under a millisecond; callers include the Silero VAD inference loop, STT plugin `recognize` paths, and TTS `ChunkedStream.collect()`.
diff --git a/agents/src/utils.test.ts b/agents/src/utils.test.ts
index 0e4657baf..b5816fb3a 100644
--- a/agents/src/utils.test.ts
+++ b/agents/src/utils.test.ts
@@ -13,6 +13,7 @@ import {
   dedent,
   delay,
   isPending,
+  mergeFrames,
   resampleStream,
   toStream,
 } from '../src/utils.js';
@@ -933,4 +934,43 @@ world
       expect(outputFrames).toEqual([]);
     });
   });
+
+  describe('mergeFrames', () => {
+    const frame = (samples: number[], sampleRate = 16000, channels = 1): AudioFrame =>
+      new AudioFrame(new Int16Array(samples), sampleRate, channels, samples.length / channels);
+
+    it('returns a non-array frame as-is', () => {
+      const f = frame([1, 2, 3]);
+      expect(mergeFrames(f)).toBe(f);
+    });
+
+    it('concatenates frame data in order', () => {
+      const merged = mergeFrames([frame([1, 2]), frame([3]), frame([4, 5, 6])]);
+      expect(Array.from(merged.data)).toEqual([1, 2, 3, 4, 5, 6]);
+      expect(merged.samplesPerChannel).toBe(6);
+      expect(merged.sampleRate).toBe(16000);
+      expect(merged.channels).toBe(1);
+    });
+
+    it('handles multi-channel frames', () => {
+      const merged = mergeFrames([frame([1, 2, 3, 4], 48000, 2), frame([5, 6], 48000, 2)]);
+      expect(Array.from(merged.data)).toEqual([1, 2, 3, 4, 5, 6]);
+      expect(merged.samplesPerChannel).toBe(3);
+      expect(merged.channels).toBe(2);
+    });
+
+    it('throws on empty buffer', () => {
+      expect(() => mergeFrames([])).toThrow(TypeError);
+    });
+
+    it('throws on sample rate mismatch', () => {
+      expect(() => mergeFrames([frame([1]), frame([2], 48000)])).toThrow('sample rate mismatch');
+    });
+
+    it('throws on channel count mismatch', () => {
+      expect(() => mergeFrames([frame([1, 2], 16000, 1), frame([3, 4], 16000, 2)])).toThrow(
+        'channel count mismatch',
+      );
+    });
+  });
 });
diff --git a/agents/src/utils.ts b/agents/src/utils.ts
index 85e351e9c..6430b9563 100644
--- a/agents/src/utils.ts
+++ b/agents/src/utils.ts
@@ -70,7 +70,7 @@ export const mergeFrames = (buffer: AudioBuffer): AudioFrame => {
     const sampleRate = buffer[0]!.sampleRate;
     const channels = buffer[0]!.channels;
     let samplesPerChannel = 0;
-    let data = new Int16Array();
+    let dataLength = 0;
 
     for (const frame of buffer) {
       if (frame.sampleRate !== sampleRate) {
@@ -81,10 +81,17 @@ export const mergeFrames = (buffer: AudioBuffer): AudioFrame => {
         throw new TypeError('channel count mismatch');
       }
 
-      data = new Int16Array([...data, ...frame.data]);
+      dataLength += frame.data.length;
       samplesPerChannel += frame.samplesPerChannel;
     }
 
+    const data = new Int16Array(dataLength);
+    let offset = 0;
+    for (const frame of buffer) {
+      data.set(frame.data, offset);
+      offset += frame.data.length;
+    }
+
     return new AudioFrame(data, sampleRate, channels, samplesPerChannel);
   }