Fix Seed-VC integration: use SeedVCWrapper instead of non-existent load_model

- Replace broken import from inference import load_model with SeedVCWrapper
- SeedVCWrapper properly loads all models (DiT, Whisper, BigVGAN, CAMPPlus, RMVPE)
- Add support for F0 conditioning and voice conversion parameters
- Implement proper voice conversion using wrapper.convert_voice()
- Keep fallback to passthrough mode if wrapper fails to load

Fixes: Could not import Seed-VC modules: cannot import name 'load_model' from 'inference'

app/services/seed_vc_service.py

@@ -17,7 +17,7 @@ class SeedVCService:
     def __init__(self, model_name: str = "seed-uvit-whisper-small-wavenet", repo_path: str = "./Seed-VC"):
         self.model_name = model_name
         self.repo_path = repo_path
-        self.model = None
+        self.wrapper = None
         self.device = None
         self.initialized = False
 
@@ -49,29 +49,34 @@ class SeedVCService:
             self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
             logger.info(f"Using device: {self.device}")
 
-            # Import Seed-VC modules
+            # Import Seed-VC wrapper
             try:
-                from inference import load_model, inference
+                from seed_vc_wrapper import SeedVCWrapper
-                self.inference_fn = inference
+                self.wrapper = SeedVCWrapper(device=self.device)
-                self.model = load_model(self.model_name, self.device)
+                logger.info("Seed-VC wrapper loaded successfully")
-                logger.info(f"Model {self.model_name} loaded successfully")
             except ImportError as e:
-                logger.warning(f"Could not import Seed-VC modules: {e}")
+                logger.warning(f"Could not import SeedVCWrapper: {e}")
                 logger.warning("Voice conversion will use passthrough mode")
-                self.model = None
+                self.wrapper = None
 
             self.initialized = True
 
         except Exception as e:
             logger.error(f"Failed to initialize Seed-VC: {e}")
             self.initialized = True  # Mark as initialized to avoid retry
-            self.model = None
+            self.wrapper = None
 
     def convert_voice(
         self,
         source_audio_path: str,
         reference_audio_path: str,
-        output_path: str
+        output_path: str,
+        diffusion_steps: int = 10,
+        length_adjust: float = 1.0,
+        inference_cfg_rate: float = 0.7,
+        f0_condition: bool = False,
+        auto_f0_adjust: bool = True,
+        pitch_shift: int = 0
     ) -> str:
         """
         Convert voice from source to match reference voice.
@@ -80,6 +85,12 @@ class SeedVCService:
             source_audio_path: Path to source audio (TTS output)
             reference_audio_path: Path to reference voice audio
             output_path: Path to save converted audio
+            diffusion_steps: Number of diffusion steps (default: 10)
+            length_adjust: Length adjustment factor (default: 1.0)
+            inference_cfg_rate: Inference CFG rate (default: 0.7)
+            f0_condition: Whether to use F0 conditioning (default: False)
+            auto_f0_adjust: Whether to automatically adjust F0 (default: True)
+            pitch_shift: Pitch shift in semitones (default: 0)
 
         Returns:
             Path to converted audio file
@@ -89,32 +100,39 @@ class SeedVCService:
         logger.info(f"Converting voice: {source_audio_path} -> {output_path}")
         logger.info(f"Using reference: {reference_audio_path}")
 
-        if self.model is None:
+        if self.wrapper is None:
             # Fallback: convert MP3 to WAV
             logger.warning("Seed-VC not available, using passthrough mode")
             return self._passthrough_convert(source_audio_path, output_path)
 
         try:
-            # Load audio files
+            # Perform voice conversion using the wrapper
-            source_audio, sr = librosa.load(source_audio_path, sr=22050)
+            # Returns numpy array of audio samples
-            reference_audio, _ = librosa.load(reference_audio_path, sr=22050)
+            audio_array = self.wrapper.convert_voice(
-            
+                source=source_audio_path,
-            # Perform voice conversion
+                target=reference_audio_path,
-            converted_audio = self.inference_fn(
+                diffusion_steps=diffusion_steps,
-                self.model,
+                length_adjust=length_adjust,
-                source_audio,
+                inference_cfg_rate=inference_cfg_rate,
-                reference_audio,
+                f0_condition=f0_condition,
-                sr=22050
+                auto_f0_adjust=auto_f0_adjust,
+                pitch_shift=pitch_shift,
+                stream_output=False  # We want the full audio array, not streaming
             )
 
+            # Determine sample rate based on f0_condition
+            sr = 44100 if f0_condition else 22050
+
             # Save output
-            sf.write(output_path, converted_audio, 22050)
+            sf.write(output_path, audio_array, sr)
             logger.info(f"Voice conversion complete: {output_path}")
 
             return output_path
 
         except Exception as e:
             logger.error(f"Voice conversion failed: {e}")
+            import traceback
+            logger.error(traceback.format_exc())
             # Fallback: convert MP3 to WAV
             logger.warning("Fallback to MP3-to-WAV conversion")
             return self._passthrough_convert(source_audio_path, output_path)