diff --git a/src/main/runtime/rpc/methods/speech.test.ts b/src/main/runtime/rpc/methods/speech.test.ts
index db730ae27..4f31ee386 100644
--- a/src/main/runtime/rpc/methods/speech.test.ts
+++ b/src/main/runtime/rpc/methods/speech.test.ts
@@ -52,4 +52,23 @@ describe('speech RPC methods', () => {
     expect(response).toMatchObject({ ok: false })
     expect(runtime.feedMobileDictation).not.toHaveBeenCalled()
   })
+
+  it('rejects oversized dictation chunks before decoding audio', async () => {
+    const runtime = {
+      getRuntimeId: () => 'test-runtime',
+      feedMobileDictation: vi.fn().mockReturnValue({ dictationId: 'dict-1' })
+    } as unknown as OrcaRuntimeService
+    const dispatcher = new RpcDispatcher({ runtime, methods: SPEECH_METHODS })
+
+    const response = await dispatcher.dispatch(
+      makeRequest('speech.dictation.chunk', {
+        dictationId: 'dict-1',
+        audioBase64: 'A'.repeat(Math.ceil((16_000 * 2 * 5) / 3) * 4 + 1),
+        sampleRate: 16_000
+      })
+    )
+
+    expect(response).toMatchObject({ ok: false })
+    expect(runtime.feedMobileDictation).not.toHaveBeenCalled()
+  })
 })
diff --git a/src/main/runtime/rpc/methods/speech.ts b/src/main/runtime/rpc/methods/speech.ts
index 71a760125..8e1ceafd5 100644
--- a/src/main/runtime/rpc/methods/speech.ts
+++ b/src/main/runtime/rpc/methods/speech.ts
@@ -3,6 +3,12 @@ import { defineMethod, type RpcMethod } from '../core'
 import { OptionalString, requiredString } from '../schemas'
 
 const AUDIO_BASE64_PATTERN = /^[A-Za-z0-9+/]*={0,2}$/
+const DICTATION_SAMPLE_RATE = 16_000
+const PCM_BYTES_PER_SAMPLE = 2
+const MAX_DICTATION_AUDIO_SECONDS = 5
+const MAX_DICTATION_AUDIO_CHUNK_BYTES =
+  DICTATION_SAMPLE_RATE * PCM_BYTES_PER_SAMPLE * MAX_DICTATION_AUDIO_SECONDS
+const MAX_DICTATION_AUDIO_CHUNK_BASE64_LENGTH = Math.ceil(MAX_DICTATION_AUDIO_CHUNK_BYTES / 3) * 4
 
 function isValidAudioBase64(value: string): boolean {
   return value.length % 4 !== 1 && AUDIO_BASE64_PATTERN.test(value)
@@ -16,6 +22,12 @@ const DictationStart = z.object({
 const DictationChunk = z.object({
   dictationId: requiredString('Missing dictation ID'),
   audioBase64: requiredString('Missing audio chunk')
+    // Why: feedMobileDictation decodes into Buffer + Float32Array; reject
+    // oversized chunks before allocation. This mirrors the mobile pending-audio budget.
+    .refine(
+      (value) => value.length <= MAX_DICTATION_AUDIO_CHUNK_BASE64_LENGTH,
+      'Audio chunk is too large'
+    )
     // Why: Buffer.from(..., 'base64') silently drops malformed bytes; reject
     // bad mobile audio chunks instead of feeding empty/corrupt PCM.
     .refine(isValidAudioBase64, 'Audio chunk must be base64'),