151 lines
5.5 KiB
Python
Executable File
151 lines
5.5 KiB
Python
Executable File
"""
|
|
Bandit-v2 separator - Multi-language/multi-stem separation
|
|
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 BANDIT_MODEL_DIR, CPU_SEPARATION
|
|
|
|
|
|
# Model configurations
|
|
BANDIT_MODELS = {
|
|
"bandit_v2_multi": {
|
|
"model_path": "checkpoint-multi_state_dict.ckpt",
|
|
"config_path": "config_dnr_bandit_v2_mus64.yaml",
|
|
"hf_repo": "miercolesv/MVSEP-Central-Backup",
|
|
"hf_subfolder": "Bandit-v2",
|
|
"stems": ["vocals", "drums", "bass", "other"],
|
|
"language": "multi",
|
|
},
|
|
"bandit_v2_eng": {
|
|
"model_path": "checkpoint-eng_state_dict.ckpt",
|
|
"config_path": "config_dnr_bandit_v2_mus64.yaml",
|
|
"hf_repo": "miercolesv/MVSEP-Central-Backup",
|
|
"hf_subfolder": "Bandit-v2",
|
|
"stems": ["vocals", "drums", "bass", "other"],
|
|
"language": "english",
|
|
},
|
|
# Additional language models available:
|
|
# - checkpoint-cmn_state_dict.ckpt (Chinese)
|
|
# - checkpoint-deu_state_dict.ckpt (German)
|
|
# - checkpoint-fao_state_dict.ckpt (Faroese)
|
|
# - checkpoint-fra_state_dict.ckpt (French)
|
|
# - checkpoint-spa_state_dict.ckpt (Spanish)
|
|
}
|
|
|
|
|
|
class BanditSeparator:
|
|
"""
|
|
Bandit-v2 separator for multi-language/multi-stem separation.
|
|
|
|
Bandit models are trained on different languages and provide
|
|
high-quality separation optimized for specific language characteristics.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
model_id: str = "bandit_v2_multi",
|
|
cpu_separation: bool = CPU_SEPARATION,
|
|
output_format: str = "mp3",
|
|
mp3_bitrate: str = "256k",
|
|
):
|
|
if model_id not in BANDIT_MODELS:
|
|
raise ValueError(f"Unknown Bandit model: {model_id}")
|
|
|
|
self.model_id = model_id
|
|
self.model_config = BANDIT_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 = BANDIT_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 Bandit 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 Bandit 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 audio into stems using Bandit-v2.
|
|
|
|
Returns:
|
|
Dict mapping stem names to output file paths
|
|
"""
|
|
output_path = Path(output_dir)
|
|
output_path.mkdir(parents=True, exist_ok=True)
|
|
|
|
try:
|
|
from audio_separator.separator import Separator
|
|
|
|
sep = Separator(
|
|
model_file_dir=str(BANDIT_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,
|
|
)
|
|
|
|
sep.load_model()
|
|
result = sep.separate(input_path)
|
|
|
|
stems = self.model_config.get("stems", ["vocals", "drums", "bass", "other"])
|
|
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"vocals.{self.output_format}"
|
|
Path(result).rename(final_path)
|
|
output_paths["vocals"] = str(final_path)
|
|
|
|
return output_paths
|
|
|
|
except ImportError:
|
|
raise ImportError(
|
|
"audio-separator is required for Bandit separation. "
|
|
"Install with: pip install audio-separator"
|
|
)
|