Postprod improvements: denoise, phone EQ, ad muting, ducking, voice mappings

- Add host mic noise reduction (afftdn + anlmdn)
- Add phone EQ bandpass on caller stem
- Mute music during ads with 2s lookahead/tail
- Increase ducking release to 3s to reduce pumping
- Add Inworld voice mappings for all regular callers
- Recording toggle endpoint, stem sync fixes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/services/audio.py

@@ -361,10 +361,6 @@ class AudioService:
             # Apply fade to prevent clicks
             audio = self._apply_fade(audio, device_sr)
 
-            # Stem recording: caller TTS
-            if self.stem_recorder:
-                self.stem_recorder.write_sporadic("caller", audio.copy(), device_sr)
-
             # Create multi-channel output with audio only on target channel
             multi_ch = np.zeros((len(audio), num_channels), dtype=np.float32)
             multi_ch[:, channel_idx] = audio
@@ -384,6 +380,9 @@ class AudioService:
                 while pos < len(multi_ch) and not self._caller_stop_event.is_set():
                     end = min(pos + chunk_size, len(multi_ch))
                     stream.write(multi_ch[pos:end])
+                    # Record each chunk as it plays so hangups cut the stem too
+                    if self.stem_recorder:
+                        self.stem_recorder.write_sporadic("caller", audio[pos:end].copy(), device_sr)
                     pos = end
 
             if self._caller_stop_event.is_set():
@@ -752,7 +751,7 @@ class AudioService:
                 mono_out = (old_samples * fade_out + new_samples * fade_in) * self._music_volume
                 outdata[:, channel_idx] = mono_out
                 if self.stem_recorder:
-                    self.stem_recorder.write("music", mono_out.copy(), device_sr)
+                    self.stem_recorder.write_sporadic("music", mono_out.copy(), device_sr)
                 self._crossfade_progress = end_progress
 
                 if self._crossfade_progress >= 1.0:
@@ -763,7 +762,7 @@ class AudioService:
                 mono_out = new_samples * self._music_volume
                 outdata[:, channel_idx] = mono_out
                 if self.stem_recorder:
-                    self.stem_recorder.write("music", mono_out.copy(), device_sr)
+                    self.stem_recorder.write_sporadic("music", mono_out.copy(), device_sr)
 
         try:
             self._music_stream = sd.OutputStream(
@@ -873,7 +872,7 @@ class AudioService:
                 chunk = self._ad_resampled[self._ad_position:self._ad_position + frames]
                 outdata[:, channel_idx] = chunk
                 if self.stem_recorder:
-                    self.stem_recorder.write("ads", chunk.copy(), device_sr)
+                    self.stem_recorder.write_sporadic("ads", chunk.copy(), device_sr)
                 self._ad_position += frames
             else:
                 if remaining > 0:

backend/services/tts.py

@@ -86,18 +86,28 @@ DEFAULT_VITS_SPEAKER = "p225"
 # Dennis, Dominus, Edward, Elizabeth, Hades, Hana, Julia, Luna, Mark, Olivia,
 # Pixie, Priya, Ronald, Sarah, Shaun, Theodore, Timothy, Wendy
 INWORLD_VOICES = {
-    # Male voices - each caller gets a unique voice matching their personality
+    # Original voice IDs
     "VR6AewLTigWG4xSOukaG": "Edward",    # Tony - fast-talking, emphatic, streetwise
     "TxGEqnHWrfWFTfGW9XjX": "Shaun",     # Rick - friendly, dynamic, conversational
     "pNInz6obpgDQGcFmaJgB": "Alex",      # Dennis - energetic, expressive, mildly nasal
     "ODq5zmih8GrVes37Dizd": "Craig",     # Earl - older British, refined, articulate
-    "IKne3meq5aSn9XLyUdCD": "Timothy",   # Marcus - lively, upbeat American
-    # Female voices - each caller gets a unique voice matching their personality
+    "IKne3meq5aSn9XLyUdCD": "Timothy",   # Marcus/Jerome - lively, upbeat American
     "jBpfuIE2acCO8z3wKNLl": "Hana",      # Jasmine - bright, expressive young female
     "EXAVITQu4vr4xnSDxMaL": "Ashley",    # Megan - warm, natural female
     "21m00Tcm4TlvDq8ikWAM": "Wendy",     # Tanya - posh, middle-aged British
     "XB0fDUnXU5powFXDhCwa": "Sarah",     # Carla - fast-talking, questioning tone
-    "pFZP5JQG7iQjIQuC4Bku": "Deborah",   # Brenda - gentle, elegant
+    "pFZP5JQG7iQjIQuC4Bku": "Deborah",   # Brenda (original) - gentle, elegant
+    # Regular caller voice IDs (backfilled)
+    "onwK4e9ZLuTAKqWW03F9": "Ronald",    # Bobby - repo man
+    "FGY2WhTYpPnrIDTdsKH5": "Julia",     # Carla (regular) - Jersey mom
+    "CwhRBWXzGAHq8TQ4Fs17": "Mark",      # Leon - male caller
+    "SOYHLrjzK2X1ezoPC6cr": "Carter",    # Carl - male caller
+    "N2lVS1w4EtoT3dr4eOWO": "Clive",     # Reggie - male caller
+    "hpp4J3VqNfWAUOO0d1Us": "Olivia",    # Brenda (regular) - ambulance driver
+    "nPczCjzI2devNBz1zQrb": "Theodore",  # Keith - male caller
+    "JBFqnCBsd6RMkjVDRZzb": "Blake",     # Andre - male caller
+    "TX3LPaxmHKxFdv7VOQHJ": "Dennis",    # Rick (regular) - male caller
+    "cgSgspJ2msm6clMCkdW9": "Priya",     # Megan (regular) - female caller
 }
 DEFAULT_INWORLD_VOICE = "Dennis"