Post-HNR RAFT-Edit Ablation Listening Study

Five ablations have complete matched audio: full hybrid, no ordinal bins, no matching discriminator, no adversarial critics, and no sparse-axis curriculum. Age and perceived gender-presentation values are model estimates; pitch and HNR are acoustic proxies.

Single-axis comparison

Age

Matched source across all five ablations. The waveform metric is age-predictor shift.

Source referenceGLOBE · young 20s / male

Ablation	-2.0	-1.0	0.0	+1.0	+2.0
Full hybridAll continuous and ordinal conditioning with realism, matching, and interpolation critics.	-2.0 Ranker Δ -0.717 Age-predictor shift -0.298 Source cosine 0.686	-1.0 Ranker Δ -0.717 Age-predictor shift -0.262 Source cosine 0.730	0.0 Ranker Δ +0.045 Age-predictor shift +0.000 Source cosine 0.911	+1.0 Ranker Δ +1.123 Age-predictor shift +0.136 Source cosine 0.661	+2.0 Ranker Δ +1.283 Age-predictor shift +0.232 Source cosine 0.572
Without ordinal binsContinuous conditioning only; removes ordinal bin structure.	-2.0 Ranker Δ -0.717 Age-predictor shift -0.261 Source cosine 0.661	-1.0 Ranker Δ -0.717 Age-predictor shift -0.222 Source cosine 0.738	0.0 Ranker Δ +0.093 Age-predictor shift +0.000 Source cosine 0.856	+1.0 Ranker Δ +1.270 Age-predictor shift +0.173 Source cosine 0.670	+2.0 Ranker Δ +1.283 Age-predictor shift +0.189 Source cosine 0.626
Without matching discriminatorRetains realism/interpolation critics but removes relative condition matching.	-2.0 Ranker Δ -0.717 Age-predictor shift -0.236 Source cosine 0.862	-1.0 Ranker Δ -0.717 Age-predictor shift -0.146 Source cosine 0.924	0.0 Ranker Δ -0.053 Age-predictor shift +0.000 Source cosine 1.000	+1.0 Ranker Δ +1.030 Age-predictor shift +0.156 Source cosine 0.887	+2.0 Ranker Δ +1.283 Age-predictor shift +0.287 Source cosine 0.806
Without adversarial criticsUses flow, target, ranker, cycle, self, and geometry objectives only.	-2.0 Ranker Δ -0.717 Age-predictor shift -0.228 Source cosine 0.858	-1.0 Ranker Δ -0.717 Age-predictor shift -0.146 Source cosine 0.931	0.0 Ranker Δ +0.003 Age-predictor shift +0.000 Source cosine 1.000	+1.0 Ranker Δ +1.041 Age-predictor shift +0.156 Source cosine 0.894	+2.0 Ranker Δ +1.283 Age-predictor shift +0.300 Source cosine 0.796
Without sparse-axis curriculumRemoves auxiliary partial-control supervision.	-2.0 Ranker Δ -0.717 Age-predictor shift -0.124 Source cosine 0.745	-1.0 Ranker Δ -0.717 Age-predictor shift -0.145 Source cosine 0.719	0.0 Ranker Δ +0.383 Age-predictor shift +0.000 Source cosine 0.776	+1.0 Ranker Δ +1.283 Age-predictor shift +0.201 Source cosine 0.661	+2.0 Ranker Δ +1.283 Age-predictor shift +0.265 Source cosine 0.666

Source ID: GLOBE::train::S_007523::00435966_000002.v2.vad

Single-axis comparison

Perceived gender presentation

Matched source across all five ablations. The waveform metric is male-probability shift.

Source referenceLibriTTS · senior 60s / female

Ablation	-2.0	-1.0	0.0	+1.0	+2.0
Full hybridAll continuous and ordinal conditioning with realism, matching, and interpolation critics.	-2.0 Ranker Δ -0.425 Male-probability shift -0.003 Source cosine 0.664	-1.0 Ranker Δ -0.425 Male-probability shift -0.003 Source cosine 0.634	0.0 Ranker Δ -0.008 Male-probability shift +0.000 Source cosine 0.998	+1.0 Ranker Δ +1.361 Male-probability shift +0.961 Source cosine 0.670	+2.0 Ranker Δ +1.575 Male-probability shift +0.983 Source cosine 0.681
Without ordinal binsContinuous conditioning only; removes ordinal bin structure.	-2.0 Ranker Δ -0.425 Male-probability shift -0.000 Source cosine 0.593	-1.0 Ranker Δ -0.425 Male-probability shift -0.000 Source cosine 0.593	0.0 Ranker Δ +0.097 Male-probability shift +0.000 Source cosine 0.903	+1.0 Ranker Δ +1.203 Male-probability shift +0.968 Source cosine 0.706	+2.0 Ranker Δ +1.575 Male-probability shift +0.990 Source cosine 0.698
Without matching discriminatorRetains realism/interpolation critics but removes relative condition matching.	-2.0 Ranker Δ -0.425 Male-probability shift -0.003 Source cosine 0.848	-1.0 Ranker Δ -0.425 Male-probability shift -0.003 Source cosine 0.955	0.0 Ranker Δ +0.033 Male-probability shift +0.000 Source cosine 1.000	+1.0 Ranker Δ +1.062 Male-probability shift +0.920 Source cosine 0.983	+2.0 Ranker Δ +1.575 Male-probability shift +0.990 Source cosine 0.917
Without adversarial criticsUses flow, target, ranker, cycle, self, and geometry objectives only.	-2.0 Ranker Δ -0.425 Male-probability shift -0.004 Source cosine 0.852	-1.0 Ranker Δ -0.425 Male-probability shift -0.004 Source cosine 0.948	0.0 Ranker Δ +0.030 Male-probability shift +0.000 Source cosine 1.000	+1.0 Ranker Δ +1.000 Male-probability shift +0.878 Source cosine 0.971	+2.0 Ranker Δ +1.575 Male-probability shift +0.989 Source cosine 0.910
Without sparse-axis curriculumRemoves auxiliary partial-control supervision.	-2.0 Ranker Δ -0.425 Male-probability shift -0.001 Source cosine 0.640	-1.0 Ranker Δ -0.425 Male-probability shift -0.001 Source cosine 0.640	0.0 Ranker Δ +0.097 Male-probability shift +0.000 Source cosine 0.719	+1.0 Ranker Δ +1.344 Male-probability shift +0.975 Source cosine 0.671	+2.0 Ranker Δ +1.575 Male-probability shift +0.981 Source cosine 0.702

Source ID: LibriTTS::4234::4234_187735_000005_000000

Single-axis comparison

Habitual pitch

Matched source across all five ablations. The waveform metric is median log-f0 shift.

Source referenceLibriTTS · senior 60s / male

Ablation	-2.0	-1.0	0.0	+1.0	+2.0
Full hybridAll continuous and ordinal conditioning with realism, matching, and interpolation critics.	-2.0 Ranker Δ -0.851 Median log-F0 shift -0.413 Source cosine 0.549	-1.0 Ranker Δ -0.851 Median log-F0 shift -0.429 Source cosine 0.611	0.0 Ranker Δ +0.024 Median log-F0 shift +0.000 Source cosine 0.925	+1.0 Ranker Δ +1.035 Median log-F0 shift +0.473 Source cosine 0.556	+2.0 Ranker Δ +1.149 Median log-F0 shift +0.927 Source cosine 0.692
Without ordinal binsContinuous conditioning only; removes ordinal bin structure.	-2.0 Ranker Δ -0.851 Median log-F0 shift -0.400 Source cosine 0.593	-1.0 Ranker Δ -0.851 Median log-F0 shift -0.338 Source cosine 0.585	0.0 Ranker Δ -0.090 Median log-F0 shift +0.000 Source cosine 0.773	+1.0 Ranker Δ +1.018 Median log-F0 shift +0.541 Source cosine 0.591	+2.0 Ranker Δ +1.149 Median log-F0 shift +0.790 Source cosine 0.648
Without matching discriminatorRetains realism/interpolation critics but removes relative condition matching.	-2.0 Ranker Δ -0.851 Median log-F0 shift -0.491 Source cosine 0.806	-1.0 Ranker Δ -0.851 Median log-F0 shift -0.378 Source cosine 0.838	0.0 Ranker Δ -0.005 Median log-F0 shift +0.000 Source cosine 1.000	+1.0 Ranker Δ +0.806 Median log-F0 shift +0.808 Source cosine 0.937	+2.0 Ranker Δ +1.149 Median log-F0 shift +0.947 Source cosine 0.843
Without adversarial criticsUses flow, target, ranker, cycle, self, and geometry objectives only.	-2.0 Ranker Δ -0.851 Median log-F0 shift -0.506 Source cosine 0.783	-1.0 Ranker Δ -0.851 Median log-F0 shift -0.383 Source cosine 0.833	0.0 Ranker Δ -0.009 Median log-F0 shift +0.000 Source cosine 1.000	+1.0 Ranker Δ +0.892 Median log-F0 shift +0.845 Source cosine 0.925	+2.0 Ranker Δ +1.149 Median log-F0 shift +0.964 Source cosine 0.811
Without sparse-axis curriculumRemoves auxiliary partial-control supervision.	-2.0 Ranker Δ -0.851 Median log-F0 shift -0.476 Source cosine 0.661	-1.0 Ranker Δ -0.851 Median log-F0 shift -0.442 Source cosine 0.622	0.0 Ranker Δ -0.144 Median log-F0 shift +0.000 Source cosine 0.639	+1.0 Ranker Δ +0.957 Median log-F0 shift +0.547 Source cosine 0.585	+2.0 Ranker Δ +1.149 Median log-F0 shift +0.707 Source cosine 0.637

Source ID: LibriTTS::6804::6804_79287_000013_000013

Single-axis comparison

Voice quality / HNR

Matched source across all five ablations. The waveform metric is hnr shift (db).

Source referenceLibriTTS · young 20s / male

Ablation	-2.0	-1.0	0.0	+1.0	+2.0
Full hybridAll continuous and ordinal conditioning with realism, matching, and interpolation critics.	-2.0 Ranker Δ -1.198 HNR shift (dB) -1.786 Source cosine 0.639	-1.0 Ranker Δ -1.198 HNR shift (dB) -0.481 Source cosine 0.663	0.0 Ranker Δ -0.000 HNR shift (dB) +0.000 Source cosine 0.971	+1.0 Ranker Δ +0.802 HNR shift (dB) +6.746 Source cosine 0.688	+2.0 Ranker Δ +0.802 HNR shift (dB) +11.453 Source cosine 0.481
Without ordinal binsContinuous conditioning only; removes ordinal bin structure.	-2.0 Ranker Δ -1.198 HNR shift (dB) -5.162 Source cosine 0.601	-1.0 Ranker Δ -1.198 HNR shift (dB) -5.080 Source cosine 0.698	0.0 Ranker Δ -0.119 HNR shift (dB) +0.000 Source cosine 0.865	+1.0 Ranker Δ +0.802 HNR shift (dB) +5.360 Source cosine 0.699	+2.0 Ranker Δ +0.802 HNR shift (dB) +12.203 Source cosine 0.550
Without matching discriminatorRetains realism/interpolation critics but removes relative condition matching.	-2.0 Ranker Δ -1.198 HNR shift (dB) -4.727 Source cosine 0.905	-1.0 Ranker Δ -0.930 HNR shift (dB) -1.476 Source cosine 0.939	0.0 Ranker Δ +0.009 HNR shift (dB) +0.000 Source cosine 1.000	+1.0 Ranker Δ +0.802 HNR shift (dB) +4.309 Source cosine 0.961	+2.0 Ranker Δ +0.802 HNR shift (dB) +13.134 Source cosine 0.890
Without adversarial criticsUses flow, target, ranker, cycle, self, and geometry objectives only.	-2.0 Ranker Δ -1.198 HNR shift (dB) -4.465 Source cosine 0.924	-1.0 Ranker Δ -0.885 HNR shift (dB) -1.630 Source cosine 0.932	0.0 Ranker Δ +0.009 HNR shift (dB) +0.000 Source cosine 1.000	+1.0 Ranker Δ +0.802 HNR shift (dB) +4.957 Source cosine 0.947	+2.0 Ranker Δ +0.802 HNR shift (dB) +13.311 Source cosine 0.872
Without sparse-axis curriculumRemoves auxiliary partial-control supervision.	-2.0 Ranker Δ -1.198 HNR shift (dB) -5.233 Source cosine 0.696	-1.0 Ranker Δ -1.070 HNR shift (dB) -3.485 Source cosine 0.648	0.0 Ranker Δ -0.126 HNR shift (dB) +0.000 Source cosine 0.784	+1.0 Ranker Δ +0.802 HNR shift (dB) +2.037 Source cosine 0.731	+2.0 Ranker Δ +0.802 HNR shift (dB) +11.679 Source cosine 0.699

Source ID: LibriTTS::8108::8108_274318_000012_000000