Configuration

Configuration

All configurable values are centralized in config/variables.py. Import and modify before creating the Separator instance.

Repository URLs

Variable	Default	Description
UVR_PUBLIC_REPO_URL	UVR model repo URL	Primary model download source
UVR_VIP_REPO_URL	UVR VIP repo URL	Paid subscriber models
AUDIO_SEPARATOR_REPO_URL	audio-separator repo URL	Fallback download source
UVR_MODEL_DATA_URL_PREFIX	UVR application data URL	Model parameter metadata
UVR_VR_MODEL_DATA_URL	VR model data URL	VR model hash → params mapping
UVR_MDX_MODEL_DATA_URL	MDX model data URL	MDX model hash → params mapping

Download Settings

Variable	Default	Description
MAX_DOWNLOAD_WORKERS	4	Parallel download threads
DOWNLOAD_CHUNK_SIZE	262144 (256 KB)	Download chunk size in bytes
DOWNLOAD_TIMEOUT	300	HTTP request timeout in seconds
HTTP_POOL_CONNECTIONS	10	urllib3 connection pool size
HTTP_POOL_MAXSIZE	10	Maximum pool connections

Example: Speed Up Downloads

import config.variables as cfg

cfg.MAX_DOWNLOAD_WORKERS = 8       # 8 parallel threads
cfg.DOWNLOAD_CHUNK_SIZE = 524288   # 512 KB chunks
cfg.DOWNLOAD_TIMEOUT = 600        # 10 minute timeout for slow connections

Separator Defaults

Variable	Default	Description
DEFAULT_MODEL_FILE_DIR	/tmp/audio-separator-models/	Where models are cached
DEFAULT_OUTPUT_FORMAT	WAV	Output audio format
DEFAULT_SAMPLE_RATE	44100	Output sample rate in Hz
DEFAULT_NORMALIZATION_THRESHOLD	0.9	Peak normalization (0.0–1.0)
DEFAULT_AMPLIFICATION_THRESHOLD	0.0	Min peak amplification
DEFAULT_OUTPUT_SINGLE_STEM	None	If set, output only this stem
DEFAULT_INVERT_USING_SPEC	False	Invert secondary stem via spectrogram
DEFAULT_USE_SOUNDFILE	False	Use soundfile for output
DEFAULT_USE_AUTOCAST	False	FP16 autocast on GPU
DEFAULT_USE_DIRECTML	False	DirectML for AMD/Intel GPUs
DEFAULT_CHUNK_DURATION	None	Seconds per chunk (None = no chunking)

Example: Change Output Defaults

import config.variables as cfg

cfg.DEFAULT_OUTPUT_FORMAT = "FLAC"
cfg.DEFAULT_SAMPLE_RATE = 48000
cfg.DEFAULT_NORMALIZATION_THRESHOLD = 0.95

Logging

Variable	Default	Description
DEFAULT_LOG_LEVEL	logging.INFO	Default log level
DEFAULT_LOG_FORMAT	%(asctime)s - %(levelname)s - %(module)s - %(message)s	Log message format

Architecture-Specific Defaults

MDX

Parameter	Default	Description
hop_length	1024	STFT hop length
segment_size	256	Segment size for processing
overlap	0.25	Overlap ratio (0.001–0.999)
batch_size	1	Batch size for inference
enable_denoise	False	Built-in denoising

VR

Parameter	Default	Description
batch_size	1	Batch size
window_size	512	Analysis window size
aggression	5	Vocal extraction intensity (-100 to 100)
enable_tta	False	Test-time augmentation
enable_post_process	False	Remove leftover vocal artifacts
post_process_threshold	0.2	Artifact detection threshold (0.1–0.3)
high_end_process	False	Mirror missing frequency range

Demucs

Parameter	Default	Description
segment_size	"Default"	Segment size ("Default" = auto)
shifts	2	Random shift augmentations
overlap	0.25	Segment overlap ratio
segments_enabled	True	Enable segment-based processing

MDXC / Roformer

Parameter	Default	Description
segment_size	256	Segment size for processing
override_model_segment_size	False	Override model's default
batch_size	1	Batch size for inference
overlap	8	Overlap between windows (2–50)
pitch_shift	0	Pitch shift in semitones

Example: Override Architecture Params

from separator import Separator

separator = Separator(
    mdx_params={"segment_size": 512, "overlap": 0.5, "batch_size": 4},
    vr_params={"window_size": 320, "aggression": 10},
    mdxc_params={"segment_size": 512, "overlap": 12},
)

Ensemble

Variable	Description
VALID_ENSEMBLE_ALGORITHMS	List of 11 valid algorithm names
DEFAULT_ENSEMBLE_ALGORITHM	avg_wave

Valid Ensemble Algorithms

Algorithm	Domain	Description
avg_wave	Time	Average waveforms
median_wave	Time	Median waveform (removes outliers)
min_wave	Time	Minimum amplitude
max_wave	Time	Maximum amplitude
avg_fft	Frequency	Average in frequency domain
median_fft	Frequency	Median in frequency domain
min_fft	Frequency	Minimum magnitude spectrum
max_fft	Frequency	Maximum magnitude spectrum
uvr_max_spec	Frequency	UVR max spectral magnitude
uvr_min_spec	Frequency	UVR min spectral magnitude
ensemble_wav	Time	UVR's native waveform ensemble

Stem Name Mapping

The STEM_NAME_MAP dictionary normalizes stem names across different models:

Alias	Normalized	Aliases
Vocals	Vocals	—
Instrumental	Instrumental	inst, karaoke, no_vocals
Drums	Drums	—
Bass	Bass	—
Guitar	Guitar	—
Piano	Piano	—
Other	Other	—
Lead Vocals	Lead Vocals	—
Backing Vocals	Backing Vocals	—
Synthesizer	Synthesizer	—
Strings	Strings	—
Woodwinds	Woodwinds	—
Brass	Brass	—

Helper Functions

get_repo_url(is_vip=False)

Returns the appropriate repository URL based on model type.

from config.variables import get_repo_url
url = get_repo_url(is_vip=True)  # VIP repo URL

get_mdx_yaml_url(filename)

Returns the full URL for an MDXC YAML config file.

from config.variables import get_mdx_yaml_url
url = get_mdx_yaml_url("config_mel_band_roformer_vocals_gabox.yaml")

get_fallback_url(filename)

Returns the fallback URL from audio-separator repo.

from config.variables import get_fallback_url
url = get_fallback_url("model_bs_roformer_ep_317_sdr_12.9755.ckpt")

Environment Variables

Variable	Description	Default
VSEP_MODEL_DIR	Model cache directory	/tmp/audio-separator-models/

# Use a custom model directory
export VSEP_MODEL_DIR=/mnt/data/vsep-models
python utils/cli.py song.mp3