Sadness under rescue rule S2, k=2. a pair defined purely by raw score movement. 90.0 % of clips score at or below zero on this emotion and the largest gap on its normalised axis is 0.449 (WIDER than the 0.25 step cap). This rule found 7,924 chains over 40,000 tracks; the strict rule found 0 at k=3.
S2 changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units. What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes. Full explanation →Interest has a median of 2.08 and is never zero, while Infatuation is zero on 87.7 % of clips — and the caption named an emotion whenever its raw score cleared an absolute 1.0. Interest therefore appeared in 94.8 % of captions and Sadness in almost none: the clause was reporting the scale of the head, not the emotion of the clip. An emotion is now named only when it lands in the top 10 % for that emotion, against a pooled tie-aware mid-rank ECDF over 132,833,726 utterances spanning every dataset and language. Interest now appears in 6.2 %, all 40 emotions occur, and a clip that is ordinary on all 40 says “no dominant emotion” rather than being forced to pick one (21.8 % of clips). This is the same scale the trajectory miner selects on, so the caption and the mining now refer to the same quantity: the mined target emotion is named in the final clip's caption on 73 % of chains, up from 46 %.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.09. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.09.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.716 before conversion and 0.644 after — it fell by 0.071. Neighbour-to-neighbour the worst pair went 0.716 → 0.644. (The earlier render, with segment 1 left raw, scores 0.601 here.) This chain started 0.70-0.80 — close, but under the identity threshold, a band where the conversion is close to a wash on this measure.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.551 in the original and +0.519 after conversion — 94 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.83 → 2.96 (+0.13) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.47. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.47.
On the corpus-wide percentile scale those become 0.44, 0.99 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 100–100 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.304 before conversion and 0.472 after — it rose by 0.168. Neighbour-to-neighbour the worst pair went 0.304 → 0.472. (The earlier render, with segment 1 left raw, scores 0.389 here.) This chain started below 0.50 — the segments really were different people, the band the conversion helps most: chains starting below 0.50 improve on this measure almost without exception.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.556 in the original and +0.550 after conversion — 99 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.68 → 3.09 (+0.41) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.24. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.24.
On the corpus-wide percentile scale those become 0.44, 0.99 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.741 before conversion and 0.754 after — it rose by 0.013. Neighbour-to-neighbour the worst pair went 0.741 → 0.754. (The earlier render, with segment 1 left raw, scores 0.543 here.) This chain started 0.70-0.80 — close, but under the identity threshold, a band where the conversion is close to a wash on this measure.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.560 in the original and +0.067 after conversion — 12 % of the delta retained, so a meaningful part of the trajectory was flattened.
Quality. Mean predicted overall quality across the segments went 2.45 → 2.81 (+0.36) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.02. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.02.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.695 before conversion and 0.670 after — it fell by 0.025. Neighbour-to-neighbour the worst pair went 0.695 → 0.670. (The earlier render, with segment 1 left raw, scores 0.605 here.) This chain started 0.50-0.70 — audibly different, a band where the conversion is close to a wash on this measure.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.548 in the original and +0.534 after conversion — 98 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.98 → 2.96 (-0.02) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.09. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.09.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.725 before conversion and 0.812 after — it rose by 0.086. Neighbour-to-neighbour the worst pair went 0.725 → 0.812. (The earlier render, with segment 1 left raw, scores 0.772 here.) This chain started 0.70-0.80 — close, but under the identity threshold, a band where the conversion is close to a wash on this measure.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.552 in the original and +0.503 after conversion — 91 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 3.07 → 3.24 (+0.17) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.00. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.00.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 100–100 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.862 before conversion and 0.871 after — it rose by 0.009. Neighbour-to-neighbour the worst pair went 0.862 → 0.871. (The earlier render, with segment 1 left raw, scores 0.723 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.546 in the original and +0.544 after conversion — 100 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.87 → 3.24 (+0.36) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.12. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.12.
On the corpus-wide percentile scale those become 0.44, 0.98 — a total move of +0.54.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 100–100 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.802 before conversion and 0.728 after — it fell by 0.074. Neighbour-to-neighbour the worst pair went 0.802 → 0.728. (The earlier render, with segment 1 left raw, scores 0.528 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.556 in the original and +0.514 after conversion — 92 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.46 → 2.80 (+0.34) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.01. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.01.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.913 before conversion and 0.925 after — it rose by 0.011. Neighbour-to-neighbour the worst pair went 0.913 → 0.925. (The earlier render, with segment 1 left raw, scores 0.871 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.547 in the original and +0.549 after conversion — 100 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.61 → 2.94 (+0.33) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.30. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.30.
On the corpus-wide percentile scale those become 0.43, 0.99 — a total move of +0.56.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.805 before conversion and 0.748 after — it fell by 0.056. Neighbour-to-neighbour the worst pair went 0.805 → 0.748. (The earlier render, with segment 1 left raw, scores 0.718 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
The emotional move did not survive. Re-scored end to end, Sadness moved +0.560 in the original and -0.098 after conversion — it changed direction. On this chain the corrected audio is not an improvement.
Quality. Mean predicted overall quality across the segments went 2.69 → 3.04 (+0.35) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.12. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.12.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.780 before conversion and 0.756 after — it fell by 0.024. Neighbour-to-neighbour the worst pair went 0.780 → 0.756. (The earlier render, with segment 1 left raw, scores 0.616 here.) This chain started 0.70-0.80 — close, but under the identity threshold, a band where the conversion is close to a wash on this measure.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.553 in the original and +0.446 after conversion — 81 % of the delta retained, which is most of it.
Quality. Mean predicted overall quality across the segments went 2.56 → 2.80 (+0.24) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.51, 1.67. In the first clip the scorer already found some Sadness here (0.51); by the last it is at 1.67.
On the corpus-wide percentile scale those become 0.92, 1.00 — a total move of +0.07.
That is why the strict rule cannot build this chain — but for the opposite reason to the jump case. Every clip here already carries some Sadness, so they all land in the crowded top tenth of the corpus-wide ranking, where roughly 90 % of clips scoring zero have consumed everything below. Measured that way the chain moves only 0.07 in total, far short of the 0.25 the strict rule demands — even though the raw scores clearly rise.
Re-ranked among the Sadness-bearing clips only — which is exactly what this rule does — the same clips read 0.38, 0.97, a move of +0.59, which is usable again.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.909 before conversion and 0.900 after — it fell by 0.009. Neighbour-to-neighbour the worst pair went 0.909 → 0.900. (The earlier render, with segment 1 left raw, scores 0.873 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.072 in the original and +0.043 after conversion — 59 % of the delta retained.
Quality. Mean predicted overall quality across the segments went 2.86 → 3.08 (+0.22) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.02. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.02.
On the corpus-wide percentile scale those become 0.44, 0.98 — a total move of +0.54.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.907 before conversion and 0.871 after — it fell by 0.036. Neighbour-to-neighbour the worst pair went 0.907 → 0.871. (The earlier render, with segment 1 left raw, scores 0.827 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.549 in the original and +0.000 after conversion — 0 % of the delta retained, so a meaningful part of the trajectory was flattened.
Quality. Mean predicted overall quality across the segments went 3.03 → 3.20 (+0.17) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.01, 1.21. In the first clip the scorer found no Sadness whatsoever (0.01); by the last it is at 1.21.
On the corpus-wide percentile scale those become 0.87, 0.99 — a total move of +0.12.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.479 before conversion and 0.326 after — it fell by 0.153. Neighbour-to-neighbour the worst pair went 0.479 → 0.326. (The earlier render, with segment 1 left raw, scores 0.341 here.) This chain started below 0.50 — the segments really were different people, the band the conversion helps most: chains starting below 0.50 improve on this measure almost without exception.
The emotional move did not survive. Re-scored end to end, Sadness moved +0.122 in the original and -0.536 after conversion — it changed direction. On this chain the corrected audio is not an improvement.
Quality. Mean predicted overall quality across the segments went 2.63 → 2.90 (+0.27) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.38. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.38.
On the corpus-wide percentile scale those become 0.43, 0.99 — a total move of +0.56.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 100–100 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.812 before conversion and 0.795 after — it fell by 0.017. Neighbour-to-neighbour the worst pair went 0.812 → 0.795. (The earlier render, with segment 1 left raw, scores 0.724 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.562 in the original and +0.540 after conversion — 96 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.92 → 3.09 (+0.17) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.24. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.24.
On the corpus-wide percentile scale those become 0.44, 0.99 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.166 before conversion and 0.517 after — it rose by 0.351. Neighbour-to-neighbour the worst pair went 0.166 → 0.517. (The earlier render, with segment 1 left raw, scores 0.454 here.) This chain started below 0.50 — the segments really were different people, the band the conversion helps most: chains starting below 0.50 improve on this measure almost without exception.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.104 in the original and +0.502 after conversion — 485 % of the delta retained, i.e. the move came out slightly larger after conversion than before.
Quality. Mean predicted overall quality across the segments went 2.70 → 2.86 (+0.16) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.15. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.15.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.316 before conversion and 0.474 after — it rose by 0.159. Neighbour-to-neighbour the worst pair went 0.316 → 0.474. (The earlier render, with segment 1 left raw, scores 0.458 here.) This chain started below 0.50 — the segments really were different people, the band the conversion helps most: chains starting below 0.50 improve on this measure almost without exception.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.555 in the original and +0.500 after conversion — 90 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.86 → 3.08 (+0.23) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.05. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.05.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 100–100 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.390 before conversion and 0.664 after — it rose by 0.274. Neighbour-to-neighbour the worst pair went 0.390 → 0.664. (The earlier render, with segment 1 left raw, scores 0.591 here.) This chain started below 0.50 — the segments really were different people, the band the conversion helps most: chains starting below 0.50 improve on this measure almost without exception.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.549 in the original and +0.544 after conversion — 99 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.79 → 3.07 (+0.27) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.85. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.85.
On the corpus-wide percentile scale those become 0.44, 1.00 — a total move of +0.56.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.267 before conversion and 0.748 after — it rose by 0.481. Neighbour-to-neighbour the worst pair went 0.267 → 0.748. (The earlier render, with segment 1 left raw, scores 0.706 here.) This chain started below 0.50 — the segments really were different people, the band the conversion helps most: chains starting below 0.50 improve on this measure almost without exception.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.568 in the original and +0.488 after conversion — 86 % of the delta retained, which is most of it.
Quality. Mean predicted overall quality across the segments went 2.88 → 3.05 (+0.18) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is 0.00, 1.15. In the first clip the scorer found no Sadness whatsoever (0.00); by the last it is at 1.15.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 150–150 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.712 before conversion and 0.679 after — it fell by 0.033. Neighbour-to-neighbour the worst pair went 0.712 → 0.679. (The earlier render, with segment 1 left raw, scores 0.731 here.) This chain started 0.70-0.80 — close, but under the identity threshold, a band where the conversion is close to a wash on this measure.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.555 in the original and +0.544 after conversion — 98 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.87 → 3.15 (+0.28) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.
This is not a strict-rule trajectory. It comes from rescue rule S2, which exists because the strict rule returns nothing at all for Sadness.
The raw scorer output across the chain is -0.00, 1.11. In the first clip the scorer found no Sadness whatsoever (-0.00); by the last it is at 1.11.
On the corpus-wide percentile scale those become 0.43, 0.98 — a total move of +0.55.
That is why the strict rule cannot build this chain. Roughly 90 % of the corpus scores exactly zero on Sadness, and tied values all collapse onto one point (about 0.45). The clips that genuinely carry Sadness sit above 0.85. There is simply nothing in between to step onto, so the only move available is one jump far wider than the 0.25 per-step cap.
What this rule changes: Raw-delta. Ignore the normalised scale altogether and require the chain to rise by at least 1.0 in RAW score units.
What it costs: Raw scores are uncalibrated and not comparable between emotions, so the same 1.0 means different things on different axes.
Same speaker? No similarity score is available here — the emolia clips in this chain are not covered by either speaker-embedding store. The chain therefore rests on the corpus's own speaker/track labelling, which is not the same as a measured check.
Voice consistency: these clips are separate recordings joined together, and no voice-similarity check could be run for this sample, so there is no measurement of how closely the voices match. You may hear the voice shift between segments. Voice conversion has not been applied yet in this build. A planned pass will re-render every segment onto the first segment's voice, which removes this effect entirely.
What was done to this chain. All 2 segments were re-synthesised with ChatterboxVC onto segment 1's voice — including segment 1 itself, converted with itself as the target — and then restored with SIDON. Converting the anchor too is what keeps the room and the reverb the same across the whole chain; leaving it raw put a change of acoustic at the first join. The words, timing and delivery still come from each original clip. The joins are 100–100 ms equal-power crossfades. The chain is normalised as one signal, so the loudness differences between segments are the ones the conversion produced, not a per-clip reset.
Did it unify the voice? On the 250-dimensional Speaker-wavLM-id verification embedding, the worst similarity between any segment and segment 1 was 0.834 before conversion and 0.842 after — it rose by 0.007. Neighbour-to-neighbour the worst pair went 0.834 → 0.842. (The earlier render, with segment 1 left raw, scores 0.661 here.) This chain started at or above 0.80 — already effectively one voice. Across the whole build that is the band where conversion tends to cost identity agreement rather than add it, and a hard identity cut would keep this chain without converting it at all. Judge it by ear against the original above.
Did the emotion survive? Re-scored end to end through the same emotion stack and the same corpus-percentile scale the chain was mined on, Sadness moved +0.553 in the original and +0.550 after conversion — 99 % of the delta retained, which is essentially all of it.
Quality. Mean predicted overall quality across the segments went 2.93 → 3.24 (+0.31) on the Empathic-Insight head. The corrected render is 48 kHz because SIDON outputs 48 kHz; the original is the 24 kHz source. Some of what you hear as “cleaner” is that bandwidth, not the conversion.