Fix noisy audio: Use WaveRNN vocoder instead of Griffin-Lim + add normalization

backend/app/voice_cloning.py

@@ -75,13 +75,27 @@ def synthesize(voice_path: Path, text: str, models_dir: Path, out_path: Path) ->
     mels = synthesizer.synthesize_spectrograms([text], [embed])
     mel = mels[0]
 
-    print("[VoiceCloning] Vocoding waveform...")
+    print("[VoiceCloning] Vocoding waveform with WaveRNN...")
     try:
-        waveform = synthesizer.griffin_lim(mel).astype(np.float32)
-    except Exception:
+        # Use the high-quality WaveRNN vocoder (much better than Griffin-Lim)
         waveform = vocoder_infer.infer_waveform(
             mel, normalize=True, batched=False, target=8000, overlap=800
         ).astype(np.float32)
+    except Exception as e:
+        print(f"[VoiceCloning] Vocoder failed: {e}, falling back to Griffin-Lim...")
+        waveform = synthesizer.griffin_lim(mel).astype(np.float32)
+
+    # Normalize waveform to prevent clipping and ensure good volume
+    max_abs_value = np.max(np.abs(waveform))
+    if max_abs_value > 0:
+        # Target peak level at -3dB (0.707 * max_int16)
+        target_level = 0.707
+        waveform = waveform * (target_level / max_abs_value)
+    
+    # Ensure waveform is in valid range for 16-bit audio
+    waveform = np.clip(waveform, -1.0, 1.0)
+    
+    print(f"[VoiceCloning] Waveform normalized - Max: {np.max(np.abs(waveform)):.4f}")
 
     out_path.parent.mkdir(parents=True, exist_ok=True)
     sf.write(out_path.as_posix(), waveform, syn_hp.sample_rate)