""" Apollo separator - Vocal enhancement and reverb removal Models from MVSEP-Central-Backup: https://huggingface.co/miercolesv/MVSEP-Central-Backup """ import os import torch from pathlib import Path from typing import Dict from ..config import APOLLO_MODEL_DIR, CPU_SEPARATION # Model configurations APOLLO_MODELS = { "apollo_vocals": { "model_path": "epoch=54-val_loss=-17.6221.ckpt", "config_path": "config_apollo_vocals_ep_54.yaml", "hf_repo": "miercolesv/MVSEP-Central-Backup", "hf_subfolder": "Apollo-Vocal-MSST", "stems": ["dry_vocals", "reverb"], "description": "Apollo vocal enhancement/reverb removal", }, "apollo_edm": { "model_path": "apollo_edm_by_essid.ckpt", "config_path": "apollo_edm_by_essid.yaml", "hf_repo": "miercolesv/MVSEP-Central-Backup", "hf_subfolder": "Politrees/UVR_resources/models/Apollo", "stems": ["enhanced", "residual"], "description": "Apollo EDM enhancement", }, "apollo_edm_big": { "model_path": "apollo_edm_big_by_essid.ckpt", "config_path": "apollo_edm_big_by_essid.yaml", "hf_repo": "miercolesv/MVSEP-Central-Backup", "hf_subfolder": "Politrees/UVR_resources/models/Apollo", "stems": ["enhanced", "residual"], "description": "Apollo EDM Big enhancement", }, } class ApolloSeparator: """ Apollo separator for vocal enhancement and reverb removal. Apollo models separate vocals into dry vocals and reverb, useful for cleaning up vocal recordings. """ def __init__( self, model_id: str = "apollo_vocals", cpu_separation: bool = CPU_SEPARATION, output_format: str = "mp3", mp3_bitrate: str = "256k", ): if model_id not in APOLLO_MODELS: raise ValueError(f"Unknown Apollo model: {model_id}") self.model_id = model_id self.model_config = APOLLO_MODELS[model_id] self.cpu_separation = cpu_separation self.output_format = output_format self.mp3_bitrate = mp3_bitrate self._model = None def _ensure_model(self): """Download model from HuggingFace if not present.""" model_dir = APOLLO_MODEL_DIR / self.model_id model_dir.mkdir(parents=True, exist_ok=True) model_path = model_dir / self.model_config["model_path"] config_path = model_dir / self.model_config["config_path"] if not model_path.exists() or not config_path.exists(): try: from huggingface_hub import hf_hub_download print(f"Downloading Apollo model: {self.model_id}") if not model_path.exists(): hf_hub_download( repo_id=self.model_config["hf_repo"], filename=f"{self.model_config['hf_subfolder']}/{self.model_config['model_path']}", local_dir=model_dir, local_dir_use_symlinks=False, ) if not config_path.exists(): hf_hub_download( repo_id=self.model_config["hf_repo"], filename=f"{self.model_config['hf_subfolder']}/{self.model_config['config_path']}", local_dir=model_dir, local_dir_use_symlinks=False, ) except Exception as e: raise RuntimeError(f"Failed to download Apollo model: {e}") return str(model_path), str(config_path) def _load_model(self): """Lazy-load the model.""" if self._model is None: model_path, config_path = self._ensure_model() self._model = None # Placeholder for actual model loading return self._model def separate_into_parts(self, input_path: str, output_dir: str) -> Dict[str, str]: """ Separate vocals into dry vocals and reverb. Returns: Dict mapping stem names to output file paths """ output_path = Path(output_dir) output_path.mkdir(parents=True, exist_ok=True) # For now, use audio-separator as fallback # Full implementation would load the actual Apollo model try: from audio_separator.separator import Separator sep = Separator( model_file_dir=str(APOLLO_MODEL_DIR), output_dir=str(output_path), output_format=self.output_format.upper(), output_bitrate=self.mp3_bitrate if self.output_format == "mp3" else None, ) # Use dereverb model as approximation sep.load_model() result = sep.separate(input_path) stems = self.model_config.get("stems", ["dry_vocals", "reverb"]) output_paths = {} if isinstance(result, list): for i, path in enumerate(result): if i < len(stems): stem_name = stems[i] final_path = output_path / f"{stem_name}.{self.output_format}" Path(path).rename(final_path) output_paths[stem_name] = str(final_path) else: final_path = output_path / f"dry_vocals.{self.output_format}" Path(result).rename(final_path) output_paths["dry_vocals"] = str(final_path) return output_paths except ImportError: raise ImportError( "audio-separator is required for Apollo separation. " "Install with: pip install audio-separator" )