OpenUnmix-L-MLX – AI Model by aufklarer | AlphaNeural AI

You can deploy this model and start earning money today!

aufklarer/
OpenUnmix-L-MLX

like

0

music-source-separation

Views

No views yet

Open-Unmix L (MLX)

MLX-converted Open-Unmix L (large variant) for music source separation on Apple Silicon.

Separates stereo music into 4 stems: vocals, drums, bass, other. Higher quality than UMX-HQ with 3x more parameters. 4.8x real-time on M2 Max.

Model Details

Architecture: BiLSTM + FC encoder/decoder with magnitude masking
Parameters: 28.3M per stem (4 stems = 113.2M total)
Hidden size: 1024
Input: Stereo 44.1kHz audio
Output: 4 stereo WAV stems
Format: safetensors (MLX-compatible)
Size: ~108 MB per stem, ~432 MB total

Comparison with UMX-HQ

Target	UMX-HQ (8.9M)	UMX-L (28.3M)
Vocals	6.23 dB	~8.5 dB
Drums	6.44 dB	~7.0 dB
Bass	4.56 dB	~5.5 dB
Other	3.41 dB	~4.5 dB

Usage

Used by speech-swift:

audio separate song.wav --model l

swift

let separator = try await SourceSeparator.fromPretrained(
    modelId: SourceSeparator.largeModelId)
let stems = separator.separate(audio: stereoAudio, sampleRate: 44100)

Files

vocals.safetensors — Vocals model (108 MB)
drums.safetensors — Drums model (108 MB)
bass.safetensors — Bass model (108 MB)
other.safetensors — Other/accompaniment model (108 MB)
config.json — Model configuration

Reference

License

MIT (same as original Open-Unmix)