Voice-profile corpus — what each build actually costs

Measured constants, three inference engines, eight build scenarios, against both compute budgets. Profile B (the Mediathek merge) is dropped — standard profile only.

Correction — an earlier version of this page was wrong. It carried a "pipelined decode, 18.8×" column and recommended porting to it as a 3.6× saving. That was a bad inference: the 18.8× came from a benchmark of a different workload (one repeated prompt, batch 256, no adapters, no reference conditioning), and the gap to our throughput was wrongly attributed entirely to pipelining.

The port was then written and measured on the real workload: it produced 1.01×. Decode simply is not a meaningful share of our loop — the same finding as scoring at 4.8 %. Generation dominates, so overlapping anything around it saves nothing. That column is gone and the honest measured rate (3.0–3.9× realtime) is the default below.

Read this first. Every number below rests on measurements, not estimates: 2,126 generations per GPU-hour and a 8.83 s mean clip from the 323,200-generation demo; 288 core-hours per node-hour and 4 GPUs per node read straight off Slurm accounting; the engine speeds from the SGLang-Omni benchmark; and the quality-versus-N figures from an exact order-statistic analysis of real 200-candidate pools.

The single biggest lever is not N — it is the batch size. Measured on this exact workload:

batch	× realtime	ms / clip
8 (what the demo ran)	3.92	1254
16	5.69	859
32	7.69	637
64 (now in use)	8.88	543

2.3× for a one-line change, and it needs no new engine, no new dependency and keeps LoRA adapters working. Every figure below uses batch 64 as the default.

The condition matrix — now 832 groups per voice

block	conditions	groups
Emotions	40 × {intense, moderate} × {free, contained} × {EN, DE}	320
VoiceNet dimensions	57 × 4 levels × {EN, DE}	456
Edge cases	screams, groans, shivers, laughter, whimpering, crying × {EN, DE}	28
Sports commentator	× {EN, DE}	2
Explicitness	× {EN, DE}, behind the age safety gate	2
Character clusters NEW	12 character-cluster LoRAs × {EN, DE} — each also rendered a second time through Chatterbox voice conversion, so the two can be compared later	24
total per voice		832

What N buys you

Share of the achievable best-of-200 gain, from the exact order-statistic analysis:

N	reward	speaker sim	genuineness	emotion
8	50%	64%	36%	38%
16	63%	74%	50%	50%
32	75%	83%	63%	64%
64	86%	90%	77%	77%
128	95%	96%	91%	91%
200	100%	100%	100%	100%

Speaker similarity saturates first — 74 % of the achievable gain already at N=16, 83 % at 32. Genuineness and emotion strength keep climbing all the way to 200, so they are what larger N actually buys.

On N = 1024. The candidate pools are 200 deep, so any N above 200 is extrapolated, not measured. Fitting the asymptotic form a + b·√(ln N) to the measured points (R² = 0.999 for reward, 0.997 for speaker similarity) gives reward 0.6199 → 0.6957 and speaker similarity 0.6540 → 0.6785 going from N=200 to N=1024 — a 5× cost increase for +0.076 and +0.024. S10 is included only to show that; S9 (300 × N=200) captures most of it for half the price and is the one to prefer. Note also that in S9 the 300 high-N voices are already 26 % of all clips despite being 5 % of the voices; at N=1024 they are 63 %.

The budgets

project	total core-h	remaining	% left	expires
REFORMO	201,600,000	68,709,721	34.1%	2026-12-31
LAIONIZE	41,660,000	37,569,231	90.2%	2027-04-30

REFORMO expires 2026-12-31 with 68.7 M core-hours unspent. Using it all would need ~14.5 M core-h/month against a recent run-rate of ~6.0 M. On the current trajectory roughly 40 million core-hours simply expire. Spend REFORMO first; keep LAIONIZE, which runs four months longer and is 90 % untouched, as the reserve.

Scenarios

scenario	engine	GPU-h	core-h	% REFORMO left	% LAIONIZE left
S1 6,000 × N=32 (flat) 159,744,000 clips · 391,817 h audio reward: N=32: 75%	batch 8 — as the demo ran (3.9x)	100,253	7,218,230	10.5%	19.2%
	batch 64 — measured, now in use (8.9x)	44,423	3,198,491	4.7%	8.5%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	12,858	925,792	1.3%	2.5%
S2 5,000 × N=32 + 1,000 × N=64 186,368,000 clips · 457,119 h audio reward: N=32: 75% / N=64: 86%	batch 8 — as the demo ran (3.9x)	116,912	8,417,669	12.3%	22.4%
	batch 64 — measured, now in use (8.9x)	51,777	3,727,973	5.4%	9.9%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	14,951	1,076,491	1.6%	2.9%
S3 5,000 × N=16 only 66,560,000 clips · 163,257 h audio reward: N=16: 63%	batch 8 — as the demo ran (3.9x)	41,897	3,016,596	4.4%	8.0%
	batch 64 — measured, now in use (8.9x)	18,635	1,341,705	2.0%	3.6%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	5,483	394,747	0.6%	1.1%
S4 5,000 × N=16 + 1,000 × N=64 119,808,000 clips · 293,862 h audio reward: N=16: 63% / N=64: 86%	batch 8 — as the demo ran (3.9x)	75,265	5,419,073	7.9%	14.4%
	batch 64 — measured, now in use (8.9x)	33,393	2,404,268	3.5%	6.4%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	9,719	699,744	1.0%	1.9%
S5 5,000 × N=16 + 900 × N=32 + 100 × N=128 101,171,200 clips · 248,150 h audio reward: N=16: 63% / N=32: 75% / N=128: 95%	batch 8 — as the demo ran (3.9x)	63,604	4,579,466	6.7%	12.2%
	batch 64 — measured, now in use (8.9x)	28,245	2,033,631	3.0%	5.4%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	8,254	594,255	0.9%	1.6%
S6 5,000 × N=16 + 900 × N=32 + 100 × N=250 111,321,600 clips · 273,047 h audio reward: N=16: 63% / N=32: 75% / N=250: 100%	batch 8 — as the demo ran (3.9x)	69,955	5,036,752	7.3%	13.4%
	batch 64 — measured, now in use (8.9x)	31,049	2,235,496	3.3%	6.0%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	9,052	651,709	0.9%	1.7%
S7 5,000 × N=16 + 900 × N=32 + 100 × N=500 132,121,600 clips · 324,065 h audio reward: N=16: 63% / N=32: 75% / N=500: 100%	batch 8 — as the demo ran (3.9x)	82,970	5,973,813	8.7%	15.9%
	batch 64 — measured, now in use (8.9x)	36,794	2,649,153	3.9%	7.1%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	10,687	769,442	1.1%	2.0%
S8 6,000 × N=128 (max quality) 638,976,000 clips · 1,567,266 h audio reward: N=128: 95%	batch 8 — as the demo ran (3.9x)	400,113	28,808,121	41.9%	76.7%
	batch 64 — measured, now in use (8.9x)	176,794	12,729,163	18.5%	33.9%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	50,533	3,638,368	5.3%	9.7%
S9 5,700 × N=32 + 300 × N=200 ← high-quality core 201,676,800 clips · 494,668 h audio reward: N=32: 75% / N=200: 100%	batch 8 — as the demo ran (3.9x)	126,491	9,107,346	13.3%	24.2%
	batch 64 — measured, now in use (8.9x)	56,006	4,032,425	5.9%	10.7%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	16,155	1,163,142	1.7%	3.1%
S10 5,700 × N=32 + 300 × N=1024 (for comparison) 407,347,200 clips · 999,132 h audio reward: N=32: 75% / N=1024: 100%	batch 8 — as the demo ran (3.9x)	255,181	18,373,007	26.7%	48.9%
	batch 64 — measured, now in use (8.9x)	112,815	8,122,672	11.8%	21.6%
	SGLang-Omni 31.2x — LoRA support UNVERIFIED	32,323	2,327,290	3.4%	6.2%

How to read it. Green is under 10 % of that budget's remainder, amber under 40 %, red above. Note that with the current engine several builds are uncomfortable, while with pipelined decode almost everything is affordable — the engine choice moves more than the scenario choice does.

Recommendation

Port pipelined decode first. It is the same transformers stack, LoRA adapters keep working, and it is a measured 3.6×. Everything else is downstream of this.
Then S5 (5,000×16 + 900×32 + 100×128). Broad coverage where speaker similarity is already 74 % saturated, a solid middle tier, and a 100-voice high-quality core for the work that needs it — at a fraction of a flat build.
Treat SGLang-Omni as a follow-up, not a plan. It is 6× faster in the benchmark but nothing has yet shown it can hot-swap LoRA adapters at arbitrary merge strength, which the entire condition matrix depends on. Confirm that before counting on it.
Always pack 4 GPUs per node. Slurm bills the whole node either way; one GPU per node quadruples every figure above.

Assumptions and where they are weak

The engine speeds are not apples-to-apples. The benchmark ran different clip lengths, different languages and no director instruction. The doc's own honest number is 1.47× SGLang over pipelined transformers; the 5.2× baseline is ours, measured on the real workload with scoring included. Treat the middle column as solid and the right-hand column as indicative.
Voice conversion for the character groups is costed at 2.5 s per clip for the top-3 only (~0.05 GPU-h/voice). Converting every candidate instead would add roughly 0.5 GPU-h per voice.
Quality fractions come from one voice's 200-candidate pools (832 groups). Group-to-group variance is real; treat them as central estimates.
Scoring is 4.8 % of the loop and is folded into the 2,126 gen/GPU-h figure. A faster generator makes scoring a proportionally larger share, so the 18.8×/31.2× rows are slightly optimistic — the scoring stage does not speed up.
No listener study. "Quality" throughout is the automatic sensor stack, which agrees with a human at roughly ρ ≈ +0.21 on spoken material.