music/benchmark/stem_probe.json
{
"_doc": "Benchmark question 1, answered by measurement.",
"mixdown": "build\\audio\\benchmark\\stems\\PROBE_MIX.wav",
"mix_samples": 1058400,
"extract": {
"strings": {
"path": "build\\audio\\benchmark\\stems\\PROBE_EXTRACT_STRINGS.wav",
"M2_best_correlation_vs_mix": 0.4988,
"M2_best_lag_samples": -5,
"M2_sample_aligned": true,
"energy_share_vs_mix": 1.4998,
"spectral_cosine_vs_mix": 0.9881
},
"percussion": {
"path": "build\\audio\\benchmark\\stems\\PROBE_EXTRACT_PERCUSSION.wav",
"M2_best_correlation_vs_mix": 0.4674,
"M2_best_lag_samples": -5,
"M2_sample_aligned": true,
"energy_share_vs_mix": 1.5528,
"spectral_cosine_vs_mix": 0.9881
},
"guitar": {
"path": "build\\audio\\benchmark\\stems\\PROBE_EXTRACT_GUITAR.wav",
"M2_best_correlation_vs_mix": 0.5159,
"M2_best_lag_samples": -5,
"M2_sample_aligned": true,
"energy_share_vs_mix": 1.6107,
"spectral_cosine_vs_mix": 0.9868
},
"keyboard": {
"path": "build\\audio\\benchmark\\stems\\PROBE_EXTRACT_KEYBOARD.wav",
"M2_best_correlation_vs_mix": 0.4717,
"M2_best_lag_samples": -5,
"M2_sample_aligned": true,
"energy_share_vs_mix": 1.4358,
"spectral_cosine_vs_mix": 0.9874
}
},
"lego": {
"path": "build\\audio\\benchmark\\stems\\PROBE_LEGO_WOODWINDS.wav",
"track": "woodwinds",
"M4_correlation_vs_source": 0.4722,
"M4_best_lag_samples": -5,
"M4_source_preserved": true,
"M4_residual_energy_ratio": 0.76278,
"M4_carries_new_content": true,
"M4_spectral_cosine_vs_source": 0.9758
},
"verdict": {
"extract": "NEITHER separation NOR compositional stems: the extract outputs are not distinct from one another (mean cross-stem cosine 0.9998), are spectrally near-identical to the mixdown, and carry MORE energy than it. The requested track name did not select content. Adaptive layering cannot be built on this output.",
"lego": "ADDITIVE COMPOSITIONAL LAYERING -- source preserved, new content present"
},
"extract_failures": {},
"M1_sum_residual": {
"n_stems_summed": 4,
"best_gain": 0.1037,
"residual_energy_ratio": 0.74229,
"residual_db": -1.29,
"sums_to_mixdown": false,
"reading": "A mask-based separator's stems sum back to the input at roughly -20 dB or better. A residual near 0 dB means the stems are NOT a decomposition of this mixdown."
},
"M5_cross_stem_distinctness": {
"pairs": {
"guitar|keyboard": {
"spectral_cosine": 0.9997,
"waveform_correlation": 0.9316
},
"guitar|percussion": {
"spectral_cosine": 0.9998,
"waveform_correlation": 0.9308
},
"guitar|strings": {
"spectral_cosine": 0.9998,
"waveform_correlation": 0.9394
},
"keyboard|percussion": {
"spectral_cosine": 0.9999,
"waveform_correlation": 0.934
},
"keyboard|strings": {
"spectral_cosine": 0.9999,
"waveform_correlation": 0.9362
},
"percussion|strings": {
"spectral_cosine": 0.9999,
"waveform_correlation": 0.9321
}
},
"mean_spectral_cosine": 0.9998,
"min_spectral_cosine": 0.9997,
"stems_are_distinct": false,
"reading": "Distinct instrument stems occupy different spectra and should sit well below 0.90 mean cosine. Near 1.0 means the 'stems' are the same content returned under different track names, which is neither separation nor layering -- it is the track name being ignored."
},
"M3_demucs": {
"available": true,
"stems_found": [
"bass",
"drums",
"other",
"vocals"
],
"per_stem": {
"strings": {
"demucs_class": "other",
"spectral_cosine": 0.994,
"waveform_correlation": 0.484
},
"percussion": {
"demucs_class": "drums",
"spectral_cosine": 0.7247,
"waveform_correlation": 0.2011
},
"guitar": {
"demucs_class": "other",
"spectral_cosine": 0.9933,
"waveform_correlation": 0.4988
},
"keyboard": {
"demucs_class": "other",
"spectral_cosine": 0.9937,
"waveform_correlation": 0.459
}
}
}
}