diff --git a/frontend/src/shared/voice/useVoiceDictation.ts b/frontend/src/shared/voice/useVoiceDictation.ts
index 7f6f283d..5c309926 100644
--- a/frontend/src/shared/voice/useVoiceDictation.ts
+++ b/frontend/src/shared/voice/useVoiceDictation.ts
@@ -175,19 +175,10 @@ export function useVoiceDictation() {
       const endpointer = hold ? null : createSilenceDetector(ctx.sampleRate);
       const streamRes = await window.openswarm?.voiceStreamStart?.();
       const streaming = streamRes?.ok === true;
-      // Whisper-mode: quiet speech gets a gentle, slow-moving boost (3x cap) so murmured dictation
-      // still clears the decode gates; loud input passes through untouched, and the gain glides so
-      // it can never pump. Applied to BOTH the streamed chunks and the stored clip.
-      let agcGain = 1;
+      // No AGC: boosting "quiet" speech clipped NORMAL speech into distortion and wrecked accuracy
+      // (Eric's garbled-history report). Whisper handles real levels fine; a whisper-mode boost can
+      // only come back as an opt-in with a proper peak limiter, never inline on the hot path.
       const capture = await createCaptureNode(ctx, (i16) => {
-        let sumSq = 0;
-        for (let i = 0; i < i16.length; i += 8) { const v = i16[i] / 0x8000; sumSq += v * v; }
-        const rawRms = Math.sqrt(sumSq / Math.max(1, Math.floor(i16.length / 8)));
-        const desired = rawRms > 0.004 && rawRms < 0.03 ? Math.min(3, 0.06 / rawRms) : 1;
-        agcGain = agcGain * 0.9 + desired * 0.1;
-        if (agcGain > 1.02) {
-          for (let i = 0; i < i16.length; i++) i16[i] = Math.max(-32768, Math.min(32767, Math.round(i16[i] * agcGain)));
-        }
         if (streaming) window.openswarm?.voiceStreamChunk?.(i16.buffer as ArrayBuffer);
         const data = new Float32Array(i16.length);
         for (let i = 0; i < i16.length; i++) data[i] = i16[i] / 0x8000;
diff --git a/frontend/src/shared/voice/voiceDictionary.ts b/frontend/src/shared/voice/voiceDictionary.ts
index ff0b0be0..a90d5d1f 100644
--- a/frontend/src/shared/voice/voiceDictionary.ts
+++ b/frontend/src/shared/voice/voiceDictionary.ts
@@ -2,9 +2,11 @@
 // LEARNED from their own dictations (a capitalized word that keeps showing up is a name worth
 // biasing toward). All local; main receives one merged comma list via voiceSetDictionary.
 
-const LEARNED_KEY = 'osw-dictation-learned';
+// v2: the v1 store learned junk from garbled transcripts and fed it BACK into decoding (a
+// degradation loop); new key orphans it, and the merge bar is higher.
+const LEARNED_KEY = 'osw-dictation-learned-v2';
 const LEARNED_CAP = 40;
-const MERGE_TOP = 20;
+const MERGE_TOP = 12;
 
 // Words that start sentences get capitalized for free; only mid-sentence capitals count as names.
 const NOUN_RE = /(?<![.!?]\s)(?<!^)\b([A-Z][a-zA-Z]{2,}(?:'s)?)\b/g;
@@ -24,7 +26,7 @@ function readLearned(): Record<string, number> {
 function pushMerged(): void {
   const counts = readLearned();
   const learned = Object.entries(counts)
-    .filter(([, n]) => n >= 2)
+    .filter(([w, n]) => n >= 3 && w.length >= 4)
     .sort((a, b) => b[1] - a[1])
     .slice(0, MERGE_TOP)
     .map(([w]) => w);