diff --git a/src/lotus/lotusDenoiseProcessor.ts b/src/lotus/lotusDenoiseProcessor.ts
index a8499388..9a17d581 100644
--- a/src/lotus/lotusDenoiseProcessor.ts
+++ b/src/lotus/lotusDenoiseProcessor.ts
@@ -207,10 +207,12 @@ async function loadDfnCore(config: LotusDenoiseConfig): Promise<DfnCore> {
   );
   const core = new mod.DeepFilterNet3Core({
     sampleRate: 48_000,
-    // 60, not 80: full-strength suppression is the main source of the
-    // "over-processed" character; a lower level keeps voice natural while
-    // the dry/wet floor handles the noise tail.
-    noiseReductionLevel: 60,
+    // Attenuation limit in dB. At 60 the background is muted outright
+    // (measured -61 to -64 dB), so every word drops into dead silence and
+    // listeners hear the mic "cutting out" at word ends. 25 keeps a quiet
+    // background bed and scored best on room-echo speech with fan and
+    // keyboard noise (PESQ 2.72 / 2.38 vs 2.33 / 2.04 at 60).
+    noiseReductionLevel: 25,
     assetConfig: { cdnUrl: dfnBase },
   });
   await core.initialize();