diff --git a/apps/server/src/speech/SpeechService.test.ts b/apps/server/src/speech/SpeechService.test.ts
index 3bec6b27a5148..2a4ba9752a6f4 100644
--- a/apps/server/src/speech/SpeechService.test.ts
+++ b/apps/server/src/speech/SpeechService.test.ts
@@ -106,6 +106,7 @@ const pcm = () => new Uint8Array(new Float32Array([0.25]).buffer);
 
 beforeEach(() => {
   vi.clearAllMocks();
+  native.finish.mockReset();
   native.backend = "Vulkan0";
   native.supportsInitialPrompt = false;
   readyModels.clear();
@@ -215,6 +216,58 @@ it.effect("unloads after the chosen idle period and reloads on the next recordin
   }).pipe(Effect.provide(layer)),
 );
 
+it.effect.each([
+  { cue: "um", remove: true, text: "Make it red, um, blue.", expected: "Make it red, um, blue." },
+  {
+    cue: "uh no",
+    remove: true,
+    text: "Make it red, uh no, blue.",
+    expected: "Make it red, uh no, blue.",
+  },
+  {
+    cue: "uh no",
+    remove: true,
+    text: "Uh, Make it red, UH  NO, blue. Uh nope.",
+    expected: "Make it red, UH NO, blue. nope.",
+  },
+  {
+    cue: " UM ",
+    remove: true,
+    text: "Uh, Make it red, um, blue.",
+    expected: "Make it red, um, blue.",
+  },
+  {
+    cue: undefined,
+    remove: true,
+    text: "Uh, Make it red, um, blue.",
+    expected: "Make it red, blue.",
+  },
+  {
+    cue: "um",
+    remove: false,
+    text: "Uh, Make it red, um, blue.",
+    expected: "Uh, Make it red, um, blue.",
+  },
+])(
+  "preserves only an explicit correction cue during batch and stream filler removal: $cue / $remove",
+  ({ cue, remove, text, expected }) =>
+    Effect.gen(function* () {
+      const speech = yield* SpeechService.SpeechService;
+      const options = {
+        ...DEFAULT_SPEECH_TRANSCRIPTION_OPTIONS,
+        speechLanguage: "en",
+        speechRemoveFillerWords: remove,
+        speechCustomWords: [{ term: "um", aliases: [] }],
+        ...(cue === undefined ? {} : { speechCorrectionWord: cue }),
+      };
+      native.transcribe.mockResolvedValueOnce({ text });
+      native.finish.mockResolvedValueOnce(text);
+      expect(yield* speech.transcribe(pcm(), options)).toBe(expected);
+      const stream = yield* speech.startStream(options);
+      expect(yield* stream.finish).toBe(expected);
+    }).pipe(Effect.scoped, Effect.provide(layer)),
+);
+
 it.effect("holds model ownership until the stream scope closes and preserves silence", () =>
   Effect.gen(function* () {
     const speech = yield* SpeechService.SpeechService;
diff --git a/apps/server/src/speech/SpeechService.ts b/apps/server/src/speech/SpeechService.ts
index 8b91677131948..6390e44df06a3 100644
--- a/apps/server/src/speech/SpeechService.ts
+++ b/apps/server/src/speech/SpeechService.ts
@@ -494,7 +494,9 @@ const make = Effect.gen(function* () {
     language: string | undefined,
   ) => {
     return (
-      options.speechRemoveFillerWords ? removeSpeechFillerWords(text, language) : text
+      options.speechRemoveFillerWords
+        ? removeSpeechFillerWords(text, language, options.speechCorrectionWord)
+        : text
     ).trim();
   };
 
diff --git a/apps/server/src/speech/fillerWords.ts b/apps/server/src/speech/fillerWords.ts
index 0181db1bf916c..5061341833c06 100644
--- a/apps/server/src/speech/fillerWords.ts
+++ b/apps/server/src/speech/fillerWords.ts
@@ -22,19 +22,21 @@ const LANGUAGE_FILLER_WORDS: Readonly<Record<string, readonly string[]>> = {
 
 const escapeRegex = (value: string) => value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
 
-export function removeSpeechFillerWords(text: string, language?: string): string {
+export function removeSpeechFillerWords(
+  text: string,
+  language?: string,
+  correctionCue?: string,
+): string {
   const baseLanguage = language?.split(/[-_]/, 1)[0]?.toLowerCase();
   const words = [
     ...UNIVERSAL_FILLER_WORDS,
     ...(baseLanguage ? (LANGUAGE_FILLER_WORDS[baseLanguage] ?? []) : []),
   ];
-  let filtered = text;
-  for (const word of words) {
-    const pattern = new RegExp(
-      `(?<![\\p{L}\\p{N}_])${escapeRegex(word)}(?![\\p{L}\\p{N}_])[,.]?`,
-      "giu",
-    );
-    filtered = filtered.replace(pattern, "");
-  }
+  const cuePattern = correctionCue?.trim().split(/\s+/).map(escapeRegex).join("\\s+");
+  const pattern = new RegExp(
+    `(?<![\\p{L}\\p{N}_])(?:(${cuePattern || "(?!)"})(?![\\p{L}\\p{N}_])|(?:${words.map(escapeRegex).join("|")})(?![\\p{L}\\p{N}_])[,.]?)`,
+    "giu",
+  );
+  const filtered = text.replace(pattern, (_match, cue: string | undefined) => cue ?? "");
   return filtered.replace(/\s{2,}/g, " ").trim();
 }
diff --git a/apps/web/src/speech/useEnvironmentSpeechInput.browser.test.tsx b/apps/web/src/speech/useEnvironmentSpeechInput.browser.test.tsx
index c47b0259046e8..398229f784907 100644
--- a/apps/web/src/speech/useEnvironmentSpeechInput.browser.test.tsx
+++ b/apps/web/src/speech/useEnvironmentSpeechInput.browser.test.tsx
@@ -402,19 +402,29 @@ it("cleans the transcript only after cleanup is separately enabled", async () =>
 });
 
 // The settings model test must show what the model heard, not what a provider rewrote.
-it("keeps a cleanup-free recording off the provider even when cleanup is on", async () => {
-  setup({ destination: "env-a" });
-  chooseTranscriptCleanup(true);
-  mocks.settings.speechCorrectionWord = "sorry";
-  mocks.projectWords = { "proj-1": [{ term: "Acme", aliases: ["ack me"] }] };
-  await using _screen = await render(<Probe cleanup={false} projectId="proj-1" />);
-  await act(() => mocks.voice.start());
-  await act(() => mocks.voice.stop());
-  expect(mocks.cleanupCalls).toEqual([]);
-  expect(mocks.recordingOptions.at(-1)?.speechCustomWords).toEqual(mocks.projectWords["proj-1"]);
-  expect(mocks.transcribeCalls).toEqual(["env-a"]);
-  expect(mocks.committed).toHaveBeenCalledExactlyOnceWith("dictated", expect.anything());
-});
+it.each([
+  { cleanup: false, authorized: true },
+  { cleanup: undefined, authorized: false },
+])(
+  "keeps a recording without authorized cleanup off the provider: $cleanup / $authorized",
+  async ({ cleanup, authorized }) => {
+    setup({ destination: "env-a" });
+    chooseTranscriptCleanup(true);
+    mocks.settings.speechCorrectionWord = "sorry";
+    if (!authorized) mocks.canDictate.delete(THREAD_ENV);
+    mocks.projectWords = { "proj-1": [{ term: "Acme", aliases: ["ack me"] }] };
+    await using _screen = await render(
+      <Probe {...(cleanup === undefined ? {} : { cleanup })} projectId="proj-1" />,
+    );
+    await act(() => mocks.voice.start());
+    await act(() => mocks.voice.stop());
+    expect(mocks.cleanupCalls).toEqual([]);
+    expect(mocks.recordingOptions.at(-1)?.speechCustomWords).toEqual(mocks.projectWords["proj-1"]);
+    expect(mocks.recordingOptions.at(-1)?.speechCorrectionWord).toBeUndefined();
+    expect(mocks.transcribeCalls).toEqual(["env-a"]);
+    expect(mocks.committed).toHaveBeenCalledExactlyOnceWith("dictated", expect.anything());
+  },
+);
 
 it("sends the project's vocabulary and name as hints, and the cue only with cleanup on", async () => {
   setup({ destination: "env-a" });
@@ -441,6 +451,14 @@ it("sends the project's vocabulary and name as hints, and the cue only with clea
     "effect",
   ]);
   expect(mocks.cleanupOptions.at(-1)).toMatchObject({ speechCorrectionWord: "scratch that" });
+  expect(mocks.recordingOptions.at(-1)?.speechCorrectionWord).toBe("scratch that");
+  chooseTranscriptCleanup(false);
+  await act(() => mocks.voice.start());
+  await act(() => mocks.voice.stop());
+  expect(mocks.recordingOptions.at(-1)?.speechCorrectionWord).toBeUndefined();
+  expect(mocks.recordingOptions.at(-1)?.speechCustomWords.map(({ term }) => term)).not.toContain(
+    "scratch that",
+  );
 });
 
 it("shares one status probe per destination across every mounted editor", async () => {
diff --git a/apps/web/src/speech/useEnvironmentSpeechInput.ts b/apps/web/src/speech/useEnvironmentSpeechInput.ts
index 1265577a7e0af..c14d2b176bdd4 100644
--- a/apps/web/src/speech/useEnvironmentSpeechInput.ts
+++ b/apps/web/src/speech/useEnvironmentSpeechInput.ts
@@ -240,7 +240,8 @@ export function useEnvironmentSpeechInput(input: HookInput) {
         } = latestRef.current;
         const cleanupGrant = current.cleanup === false ? null : readCleanupGrant();
         const cleanupEnabled = cleanupGrant !== null;
-        const cue = cleanupEnabled ? settings.speechCorrectionWord : "";
+        const cue =
+          cleanupEnabled && latestRef.current.cleanupPrepared ? settings.speechCorrectionWord : "";
         const speechCustomWords = resolveSpeechVocabulary({
           personal: settings.speechCustomWords,
           project,
@@ -262,6 +263,7 @@ export function useEnvironmentSpeechInput(input: HookInput) {
           speechTranslateToEnglish: settings.speechTranslateToEnglish,
           speechCustomWords,
           speechRemoveFillerWords: settings.speechRemoveFillerWords,
+          ...(cue.trim() ? { speechCorrectionWord: cue } : {}),
         };
       },
       getMicrophoneId: () => latestRef.current.clientSettings.voiceMicrophone,
diff --git a/packages/contracts/src/speech.ts b/packages/contracts/src/speech.ts
index a78f131019c74..1fc1ecaa4ceef 100644
--- a/packages/contracts/src/speech.ts
+++ b/packages/contracts/src/speech.ts
@@ -117,12 +117,17 @@ export const SpeechPostProcessingPromptSettings = Schema.Struct({
 });
 export type SpeechPostProcessingPromptSettings = typeof SpeechPostProcessingPromptSettings.Type;
 
+/** A word the person says to flag a spoken correction, such as "sorry" or "err". */
+export const SpeechCorrectionCue = Schema.String.check(Schema.isMaxLength(50));
+
 /** Recognition preferences that travel with every recording. */
 export const SpeechTranscriptionOptions = Schema.Struct({
   speechLanguage: SpeechLanguage,
   speechTranslateToEnglish: Schema.Boolean,
   speechCustomWords: SpeechCustomWords,
   speechRemoveFillerWords: Schema.Boolean,
+  /** Preserve this cue for separately consented transcript cleanup. */
+  speechCorrectionWord: Schema.optionalKey(SpeechCorrectionCue),
 });
 export type SpeechTranscriptionOptions = typeof SpeechTranscriptionOptions.Type;
 export const DEFAULT_SPEECH_TRANSCRIPTION_OPTIONS: SpeechTranscriptionOptions = {
@@ -133,9 +138,6 @@ export const DEFAULT_SPEECH_TRANSCRIPTION_OPTIONS: SpeechTranscriptionOptions =
 };
 export const SPEECH_MAX_OPTIONS_BYTES = 128_000;
 
-/** A word the person says to flag a spoken correction, such as "sorry" or "err". */
-export const SpeechCorrectionCue = Schema.String.check(Schema.isMaxLength(50));
-
 export const SpeechPostProcessingOptions = Schema.Struct({
   speechCustomWords: SpeechCustomWords,
   speechPostProcessingPrompt: SpeechPostProcessingPromptSettings,