Remove fade-in on sentence beginning

engine/core/duplex_pipeline.py

@@ -555,11 +555,9 @@ class DuplexPipeline:
                             })
                             first_audio_sent = True
 
-                        # Keep very short fade-in for non-first sentence to preserve consonant attack.
-                        fade_in_ms = 2 if spoken_sentence_count == 0 else 1
                         await self._speak_sentence(
                             sentence,
-                            fade_in_ms=fade_in_ms,
+                            fade_in_ms=0,
                             fade_out_ms=8,
                         )
                         spoken_sentence_count += 1
@@ -585,10 +583,9 @@ class DuplexPipeline:
                         )
                     })
                     first_audio_sent = True
-                fade_in_ms = 2 if spoken_sentence_count == 0 else 1
                 await self._speak_sentence(
                     remaining_text,
-                    fade_in_ms=fade_in_ms,
+                    fade_in_ms=0,
                     fade_out_ms=8,
                 )
             
@@ -683,7 +680,7 @@ class DuplexPipeline:
 
         return False
 
-    async def _speak_sentence(self, text: str, fade_in_ms: int = 2, fade_out_ms: int = 8) -> None:
+    async def _speak_sentence(self, text: str, fade_in_ms: int = 0, fade_out_ms: int = 8) -> None:
         """
         Synthesize and send a single sentence.
         
@@ -744,7 +741,7 @@ class DuplexPipeline:
         sample_rate: int,
         fade_in: bool = False,
         fade_out: bool = False,
-        fade_in_ms: int = 2,
+        fade_in_ms: int = 0,
         fade_out_ms: int = 8,
     ) -> bytes:
         """Apply short edge fades to reduce click/pop at sentence boundaries."""
@@ -756,7 +753,7 @@ class DuplexPipeline:
             if samples.size == 0:
                 return pcm_bytes
 
-            if fade_in:
+            if fade_in and fade_in_ms > 0:
                 fade_in_samples = int(sample_rate * (fade_in_ms / 1000.0))
                 fade_in_samples = max(1, min(fade_in_samples, samples.size))
                 samples[:fade_in_samples] *= np.linspace(0.0, 1.0, fade_in_samples, endpoint=True)