nubeaudio-mdx — MDX23C InstVoc HQ (Int8 CoreML)

CoreML-converted, Int8-quantized version of MDX23C-8KFFT-InstVoc_HQ.ckpt for on-device vocal/instrumental separation in NubeAudio — an iOS music player.

The .mlpackage runs the mask-prediction core only. STFT/iSTFT are computed in Swift with Accelerate/vDSP so the model stays portable and the spectrogram pipeline is debuggable end-to-end.

File

File	Size	Format
mdx23c_instvoc_hq.mlpackage.zip	~40 MB	zipped CoreML mlprogram, weights Int8

Unzipped: mdx23c_instvoc_hq.mlpackage ≈ 107 MB.

I/O contract

Input spectrogram — MLMultiArray<Float32>, shape [1, 4, 4096, T]

• 4 = 2 audio channels × [real, imag] interleaved
• 4096 = dim_f (truncated from n_fft/2 + 1 = 4097)
• T = time frames; flexible (RangeDim), nominal 256 for a 5.92 s chunk

Output stems — MLMultiArray<Float32>, shape [1, 2, 4, 4096, T]

• Outer 2 = [Vocals, Instrumental]
• Inner 4, 4096, T = same layout as input — the model returns time-domain-equivalent complex spectrograms ready for iSTFT

Spectrogram parameters (must match exactly)

Encoded in user_defined_metadata:

Param	Value
sample_rate	44100
n_fft	8192
hop_length	1024
dim_f	4096
chunk_size	261120 samples (~5.92 s)
window	Hann periodic
instruments	Vocals,Instrumental

Inputs at other sample rates must be resampled to 44.1 kHz before STFT.

Conversion source

• Architecture: KUIELab TFC-TDF v3 / MDX23C — code from ZFTurbo/Music-Source-Separation-Training.
• Original checkpoint: MDX23C-8KFFT-InstVoc_HQ.ckpt (trained by ZFTurbo, distributed via UVR).
• Conversion: torch.jit.trace on a wrapper that exposes encoder→decoder→final_conv (skipping STFT/iSTFT), then coremltools.convert(convert_to="mlprogram") + linear_quantize_weights(mode="linear_symmetric", dtype="int8").
• Verified max_abs_diff(traced, eager) == 0.0 before quantization.

Credits

• Model: ZFTurbo (architecture and weights), KUIELab (TFC-TDF v3).
• UVR project: Anjok07.
• CoreML packaging: NubeAudio.