VOICE-CONVERTED BUILD · corrected tiers · what was done & what it measured · the ORIGINAL (uncorrected) site

Emotion trajectories, voice-corrected — a parallel build

67 emotion tiers · 1,304 chains · 3,438 segments re-voiced onto their own chain's segment 1 with ChatterboxVC + SIDON. The uncorrected site is untouched and every card here carries both renders.

Read this first. These pages are an experiment with a published result, not a straight upgrade. The emotional trajectories survive the conversion almost intact — that was the risk, and it did not happen. What the conversion does to the voice is not uniform: it is a large improvement on the chains that were genuinely broken and a measurable loss on the chains that were already fine. Both renders are shipped on every card so you can judge by ear rather than take the numbers on trust. The full method and the measurements →

1,304chains converted

3,438segments re-voiced

1.5GPU-hours

0.756 → 0.781median worst-to-anchor identity cosine, processing-matched

96 %median emotion delta retained

The one-paragraph verdict. The emotion survives: re-scored end to end through the same stack and the same corpus-percentile scale the chains were mined on, the trajectory retains a median 90 % of its size in the default render and keeps its direction in 91 % of chains. The voice is a different story, and the average hides it: conversion pulls every chain towards a common level of identity agreement, so what it does depends on where the chain started. For the 25 % whose segments really were different people (cosine below 0.50) it is a large, near-universal win: 0.221 → 0.653. For the 40 % already at or above 0.80 it is roughly neutral at best (-0.015) — and those chains never needed converting. The full breakdown →

Would a hard identity cut be better than converting? For most of the corpus, yes. Of these 1,304 chains, 40 % already clear a 0.80 identity cut, and on those conversion costs 0.117 while helping only 2 % of them. Below 0.70 it is genuinely worth +0.173 — but those are also the chains where you can hear the conversion happen. Section 10 has the thresholds, the corpus-wide counts and the −id/−tbr calibration →

What it costs. Intelligibility is the clearest price: character error rate against the corpus transcript goes 10.0 % → 12.3 % on a stratified sample — the speech survives but is measurably less cleanly articulated.

What each card gives you. The original render and the corrected render side by side, both labelled; a third “uniform” render behind an expander in which segment 1 was converted too; per-chunk playback that follows whichever render you select; the chain's own before/after identity cosine and before/after emotion delta; and everything the original card had — the generated paragraph, the SCRIPT screenplay with its hoisted constants and its (burst) markers, and the full caption behind an expander.

Scope. Emotion tiers only. The VN1 VoiceNet tiers and the vprof_vc voice-profile tiers were deliberately not converted — the profiles are already one cloned voice per chain and have nothing to fix. They are all still on the original site, unchanged.

Manifest tiers — the subsets he will train on

tier	rule	chains	segments re-voiced	cos→seg1 orig	corrected	uniform	emotion retained	corrected	original
`emotion__B1__T0.20__C0.25__INTERNAL`	B1	20	50	0.752	0.678	0.753	97 %	listen →	uncorrected
`emotion__B1__T0.25__C0.20__INTERNAL`	B1	20	66	0.753	0.639	0.726	89 %	listen →	uncorrected
`emotion__B1__T0.25__C0.25__INTERNAL`	B1	20	50	0.763	0.670	0.781	93 %	listen →	uncorrected
`emotion__B1__T0.40__C0.25__INTERNAL`	B1	20	62	0.728	0.669	0.747	95 %	listen →	uncorrected
`emotion__B1__T0.50__C0.25__INTERNAL`	B1	20	73	0.731	0.625	0.707	95 %	listen →	uncorrected
`emotion__B1__T0.60__C0.25__INTERNAL`	B1	20	72	0.733	0.639	0.715	95 %	listen →	uncorrected
`emotion__B1__T0.70__C0.25__INTERNAL`	B1	20	80	0.718	0.615	0.755	96 %	listen →	uncorrected
`emotion__B1__T0.80__C0.25__INTERNAL`	B1	20	80	0.664	0.593	0.704	99 %	listen →	uncorrected
`emotion_twosided__AB2__T0.20__C0.25__INTERNAL`	AB2	20	53	0.737	0.644	0.746	89 %	listen →	uncorrected
`emotion_twosided__AB2__T0.25__C0.25__INTERNAL`	AB2	20	48	0.830	0.677	0.819	83 %	listen →	uncorrected
`merged_emo_vn__T0.20__C0.25__INTERNAL`	B1 UNION VN1	20	51	0.732	0.673	0.768	89 %	listen →	uncorrected
`proxy_spearman__PXR__T0.20__C0.25__INTERNAL`	PXR	20	44	0.774	0.701	0.802	95 %	listen →	uncorrected
`proxy_spearman__PXR__T0.25__C0.25__INTERNAL`	PXR	20	49	0.766	0.668	0.785	91 %	listen →	uncorrected
`proxy_spearman__PXR__T0.40__C0.25__INTERNAL`	PXR	20	64	0.760	0.643	0.736	90 %	listen →	uncorrected
`proxy_spearman__PXR__T0.50__C0.25__INTERNAL`	PXR	20	74	0.776	0.578	0.770	99 %	listen →	uncorrected
`proxy_spearman__PXR__T0.60__C0.25__INTERNAL`	PXR	20	77	0.672	0.591	0.685	98 %	listen →	uncorrected
`proxy_spearman__PXR__T0.70__C0.25__INTERNAL`	PXR	2	8	0.441	0.514	0.668	95 %	listen →	uncorrected
`proxy_taillift__PXR__T0.20__C0.25__INTERNAL`	PXR	20	38	0.803	0.702	0.848	84 %	listen →	uncorrected
`proxy_taillift__PXR__T0.25__C0.25__INTERNAL`	PXR	20	54	0.748	0.689	0.750	89 %	listen →	uncorrected
`proxy_taillift__PXR__T0.40__C0.25__INTERNAL`	PXR	20	61	0.767	0.642	0.757	95 %	listen →	uncorrected
`proxy_taillift__PXR__T0.50__C0.25__INTERNAL`	PXR	20	70	0.707	0.566	0.685	94 %	listen →	uncorrected
`proxy_taillift__PXR__T0.60__C0.25__INTERNAL`	PXR	20	76	0.709	0.601	0.735	98 %	listen →	uncorrected
`proxy_taillift__PXR__T0.70__C0.25__INTERNAL`	PXR	2	8	0.441	0.484	0.595	97 %	listen →	uncorrected

Rule × chain length

tier	rule	chains	segments re-voiced	cos→seg1 orig	corrected	uniform	emotion retained	corrected	original
`k-AB2-k3`	AB2	20	40	0.778	0.640	0.803	93 %	listen →	uncorrected
`k-AB2-k4`	AB2	20	60	0.755	0.703	0.782	89 %	listen →	uncorrected
`k-AB2-k5`	AB2	20	80	0.675	0.632	0.771	98 %	listen →	uncorrected
`k-B1-k2`	B1	20	20	0.841	0.723	0.854	76 %	listen →	uncorrected
`k-B1-k3`	B1	20	40	0.795	0.696	0.811	94 %	listen →	uncorrected
`k-B1-k4`	B1	20	60	0.661	0.634	0.649	92 %	listen →	uncorrected
`k-B1-k5`	B1	20	80	0.664	0.614	0.745	95 %	listen →	uncorrected
`k-PXR-k3`	PXR	20	40	0.835	0.719	0.835	92 %	listen →	uncorrected
`k-PXR-k4`	PXR	20	60	0.783	0.706	0.760	99 %	listen →	uncorrected
`k-PXR-k5`	PXR	20	80	0.730	0.694	0.813	98 %	listen →	uncorrected

One corpus at a time

tier	rule	chains	segments re-voiced	cos→seg1 orig	corrected	uniform	emotion retained	corrected	original
`c-emolia-AB2`	AB2	20	53	0.787	0.701	0.749	89 %	listen →	uncorrected
`c-emolia-B1`	B1	20	62	0.669	0.610	0.731	89 %	listen →	uncorrected
`c-emolia-PXR`	PXR	20	51	0.772	0.644	0.763	96 %	listen →	uncorrected
`c-eurospeech-AB2`	AB2	20	61	0.449	0.638	0.834	84 %	listen →	uncorrected
`c-eurospeech-B1`	B1	20	57	0.643	0.746	0.879	90 %	listen →	uncorrected
`c-eurospeech-PXR`	PXR	20	56	0.329	0.699	0.866	89 %	listen →	uncorrected
`c-mls-AB2`	AB2	20	57	0.926	0.806	0.904	96 %	listen →	uncorrected
`c-mls-B1`	B1	20	44	0.930	0.786	0.905	99 %	listen →	uncorrected
`c-mls-PXR`	PXR	20	57	0.917	0.774	0.908	98 %	listen →	uncorrected
`c-podcast-AB2`	AB2	20	50	0.701	0.658	0.805	94 %	listen →	uncorrected
`c-podcast-B1`	B1	20	53	0.584	0.565	0.737	93 %	listen →	uncorrected
`c-podcast-PXR`	PXR	20	49	0.240	0.555	0.659	97 %	listen →	uncorrected
`c-snippets-AB2`	AB2	20	51	0.162	0.435	0.565	96 %	listen →	uncorrected
`c-snippets-B1`	B1	20	61	0.179	0.440	0.517	93 %	listen →	uncorrected
`c-snippets-PXR`	PXR	20	48	0.139	0.499	0.612	99 %	listen →	uncorrected
`c-evasnippets-AB2`	AB2	20	64	0.482	0.663	0.812	97 %	listen →	uncorrected
`c-evasnippets-B1`	B1	20	59	0.769	0.659	0.846	99 %	listen →	uncorrected
`c-evasnippets-PXR`	PXR	20	59	0.863	0.765	0.888	99 %	listen →	uncorrected

Speaker-cleaned set

tier	rule	chains	segments re-voiced	cos→seg1 orig	corrected	uniform	emotion retained	corrected	original
`sc-AB2-k3`	AB2	20	40	0.844	0.733	0.833	94 %	listen →	uncorrected
`sc-AB2-k4`	AB2	20	60	0.835	0.689	0.819	94 %	listen →	uncorrected
`sc-AB2-k5`	AB2	20	80	0.825	0.729	0.817	97 %	listen →	uncorrected

Low-resource cells

tier	rule	chains	segments re-voiced	cos→seg1 orig	corrected	uniform	emotion retained	corrected	original
`rare-AB2-pairs`	AB2	20	45	0.616	0.577	0.708	96 %	listen →	uncorrected
`rare-PXR-pairs`	PXR	20	55	0.620	0.659	0.755	99 %	listen →	uncorrected
`rare-B1-pairs`	B1	20	38	0.915	0.791	0.901	99 %	listen →	uncorrected
`rare-lang`	mixed	20	20	0.910	0.745	0.860	109 %	listen →	uncorrected

Rescue rules (looser rules, kept separate)

tier	rule	chains	segments re-voiced	cos→seg1 orig	corrected	uniform	emotion retained	corrected	original
`sad-Sadness-S3-k2`	S3	20	20	0.733	0.655	0.774	95 %	listen →	uncorrected
`sad-Sadness-S3-k3`	S3	20	40	0.749	0.649	0.744	97 %	listen →	uncorrected
`sad-Sadness-S1-k3`	S1	20	40	0.688	0.676	0.776	48 %	listen →	uncorrected
`sad-Sadness-S2-k2`	S2	20	20	0.733	0.639	0.748	95 %	listen →	uncorrected
`sad-Sadness-S4-k2`	S4	20	20	0.778	0.757	0.824	95 %	listen →	uncorrected
`sad-Awe-S3-k2`	S3	20	20	0.795	0.712	0.790	99 %	listen →	uncorrected
`sad-Distress-S3-k2`	S3	20	20	0.718	0.702	0.795	93 %	listen →	uncorrected
`sad-Disappointment-S3-k2`	S3	20	20	0.754	0.763	0.818	94 %	listen →	uncorrected
`sad-Helplessness-BASE-k3`	BASE	20	40	0.753	0.722	0.797	100 %	listen →	uncorrected

Provenance

The converted audio is derived from emolia, eurospeech, mls, podcast, snippets and evasnippets. The annotations are CC-BY-4.0. This is a listening demo, not a corpus release, and the converted audio is synthetic speech: it carries the words and delivery of one recording in the voice of another.