Separação de fontes

Open-Unmix HQ divide uma faixa musical estéreo em quatro stems independentes — vocais, bateria, baixo e outros. Quatro modelos BiLSTM independentes (um por stem) produzem máscaras de magnitude sobre a STFT da mistura; um pós-filtro Wiener opcional os reconcilia. Executa no Apple Silicon via MLX.

Dois motores estão disponíveis: Open-Unmix HQ (leve, o padrão) e HTDemucs (Demucs v4) — um modelo Hybrid Transformer de maior qualidade selecionado com --engine htdemucs. Ambos executam no Apple Silicon via MLX e produzem os mesmos quatro stems a 44.1 kHz.

Visão geral

• 4 stems per track — vocals, drums, bass, other. Each is a 2-channel 44.1 kHz stem file.
• Magnitude-mask model — each stem model predicts a non-negative mask applied to the mixture spectrogram; phase is taken from the mixture.
• Wiener post-filter (optional) — soft-mask refinement across all 4 stems so they sum coherently back to the mixture. Adds ~0.5 dB SDR.
• Small footprint — 8.9M params per stem, ~136 MB total for all 4 stems.
• Apache-2.0 — upstream weights under MIT, our CoreML/MLX conversion under Apache-2.0.

Arquitetura

Four independent stems, each a copy of the same network:

Modelo

HTDemucs (Demucs v4)

For higher separation quality — especially on bass and drums — the package also ships HTDemucs, Meta's Hybrid Transformer Demucs. It merges a spectrogram branch and a waveform branch through a cross-domain transformer; the shipped htdemucs_ft variant is a bag of four fine-tuned sub-models, one per stem. Weights download from HuggingFace on first use. On a directional MUSDB-sample benchmark (museval / BSSEval v4) it averages +3.01 dB SDR over UMX-HQ, with the biggest gains on bass (+5.75 dB).

Início rápido — Swift

import SourceSeparation
import AudioCommon

let separator = try await SourceSeparator.fromPretrained()

let stereo = try AudioFileLoader.loadStereo(
    url: URL(fileURLWithPath: "song.wav"),
    targetSampleRate: 44100
)

let stems = separator.separate(audio: stereo, sampleRate: 44100)
// stems[.vocals], stems[.drums], stems[.bass], stems[.other]
// Each is [[Float]] — left channel, right channel.

try WAVWriter.writeStereo(
    left: stems[.vocals]![0],
    right: stems[.vocals]![1],
    sampleRate: 44100,
    to: URL(fileURLWithPath: "vocals.wav")
)

Pass wiener: true (default) for best quality. Pass targets: [.vocals] to extract only a subset of stems and skip the other models.

CLI

speech separate song.wav                              # all 4 stems into song_stems/ (Open-Unmix)
speech separate song.wav --engine htdemucs            # Demucs v4 — higher quality
speech separate song.wav --engine htdemucs --htdemucs-precision int8  # smaller int8 bundle
speech separate song.wav --stems vocals               # vocals only
speech separate song.wav --stems vocals,drums         # subset
speech separate song.wav --output-dir /tmp/stems/     # custom output dir
speech separate song.wav --verbose                    # show timing

Quando usar