Update denoiser.py

denoiser.py

@@ -202,37 +202,58 @@ class Denoiser:
         except Exception as e:
             logger.warning(f"[Denoiser] DeepFilterNet unavailable ({e})")
 
-        # TWO-PASS noise reduction for clearly audible noise:
-        # Pass 1 — Stationary noise (hum, hiss, fan, AC)
-        # Pass 2 — Non-stationary noise (intermittent background sounds)
+    def _remove_background_noise(self, audio, sr):
+        # ── Try DeepFilterNet (SOTA) first ──────────────────────────
+        try:
+            result = self._deepfilter(audio, sr)
+            print("[Denoiser] ✅ DeepFilterNet noise removal done")
+            return result, "DeepFilterNet"
+        except Exception as e:
+            logger.warning(f"[Denoiser] DeepFilterNet unavailable ({e})")
+
+        # ── Try RNNoise (best for voice, no musical artifacts) ───────
+        try:
+            result = self._rnnoise(audio, sr)
+            print("[Denoiser] ✅ RNNoise removal done")
+            return result, "RNNoise"
+        except Exception as e:
+            logger.warning(f"[Denoiser] RNNoise unavailable ({e})")
+
+        # ── Fallback: noisereduce with mild settings ──────────────────
+        # IMPORTANT: Keep prop_decrease LOW (0.50-0.60) to avoid musical
+        # noise artifacts. Two aggressive passes make musical noise WORSE.
         try:
             import noisereduce as nr
 
-            # Pass 1: Remove stationary background noise (hiss, hum, fan)
-            pass1 = nr.reduce_noise(
+            # Single gentle pass — avoids musical noise artifacts
+            cleaned = nr.reduce_noise(
                 y=audio, sr=sr,
-                stationary=True,
-                prop_decrease=0.95,       # 95% — very aggressive on stationary noise
-                n_std_thresh_stationary=1.2,  # lower threshold = catches more noise
-                freq_mask_smooth_hz=500,
-                time_mask_smooth_ms=50,
-            ).astype(np.float32)
-
-            # Pass 2: Remove non-stationary noise (movement, traffic, intermittent)
-            pass2 = nr.reduce_noise(
-                y=pass1, sr=sr,
-                stationary=False,
-                prop_decrease=0.85,       # 85% on non-stationary
-                freq_mask_smooth_hz=500,
-                time_mask_smooth_ms=50,
+                stationary=False,         # non-stationary handles both types
+                prop_decrease=0.55,       # gentle — avoids buzzing artifacts
+                freq_mask_smooth_hz=1000, # heavy smoothing = less musical noise
+                time_mask_smooth_ms=100,  # heavy smoothing = less musical noise
+                n_std_thresh_stationary=2.0,  # higher = less aggressive
             ).astype(np.float32)
-
-            print("[Denoiser] ✅ Two-pass noisereduce done (stationary + non-stationary)")
-            return pass2, "noisereduce (2-pass)"
+            print("[Denoiser] ✅ noisereduce (gentle, artifact-free) done")
+            return cleaned, "noisereduce"
         except Exception as e:
             logger.warning(f"noisereduce failed: {e}")
             return audio, "none"
 
+    def _rnnoise(self, audio, sr):
+        """
+        RNNoise — Mozilla's RNN-based noise suppressor.
+        Specifically trained on voice, produces zero musical artifacts.
+        Requires: pip install rnnoise-python
+        """
+        import rnnoise
+        # RNNoise requires 48kHz mono float32
+        audio_48k = self._resample(audio, sr, 48000) if sr != 48000 else audio
+        denoised  = rnnoise.reduce_noise(audio_48k, 48000)
+        # Resample back to original SR
+        result = self._resample(denoised, 48000, sr) if sr != 48000 else denoised
+        return result.astype(np.float32)
+
     def _deepfilter(self, audio, sr):
         if not self._df_loaded:
             from df.enhance import enhance, init_df