LAION · Curated reference voices · Five languages

Humaneness Podcast
Duration Comparison

The same words, voices and seeds with three timing settings: the original longer allocation,3–4 seconds per SCRIPT turn, and no requested timing. Reference pairs are chosen from the actual input recordings using Content Enjoyment. The listener-approved English pair en-real-2 is retained.

Audition selected inputsPace and qualityThree-way listening comparisonsExactly what changed
652 / 803new shorter / untimed takes
397existing baselines reused
0fully scored triplets
CFG 2.5same settings throughout

Selected reference recordings

Except for the explicitly listener-approved English pair, each language uses the pair with the highest minimum Content Enjoyment across its two identity anchors and the conversation context. A weak input cannot hide behind a good average. These are automatic EARS Medium estimates, with all selection scores saved.

Language / referenceS1 CE / inputS2 CE / inputContext CE / inputSelection
English
en-real-2		5.07
5.03
5.54
Listener explicitly approved this reference pair
German
de-notebooklm-new		5.05
4.98
5.30
Highest minimum Content Enjoyment across S1, S2 and context; mean CE breaks ties
French
fr-real-3		4.82
4.89
5.14
Highest minimum Content Enjoyment across S1, S2 and context; mean CE breaks ties
Spanish
es-real-2		4.82
4.89
5.10
Highest minimum Content Enjoyment across S1, S2 and context; mean CE breaks ties
Dutch
nl-real-2		4.93
4.95
5.38
Highest minimum Content Enjoyment across S1, S2 and context; mean CE breaks ties

Reference selection is complete.

Does shorter timing sound more natural?

The tables compare only scripts for which all three timing conditions have audio and measurements. Each triplet keeps spoken text, seed, voice references, earlier conversation, emotions, GENERAL and sampling settings fixed. Recognized words per minute measures delivery pace; WER reveals when faster output loses words. Higher pace alone is not a quality improvement.

All five languages

TimingMatched takesCE ↑Genuine ↑Identity cosine ↑WER ↓Recognized words/minActual secondsSafety-cap hits
Original longer0pendingpendingpendingpendingpendingpending0
Short3–4s/turn0pendingpendingpendingpendingpendingpending0
No duration request0pendingpendingpendingpendingpendingpending0
Per-language breakdown

English

TimingMatched takesCE ↑Genuine ↑Identity cosine ↑WER ↓Recognized words/minActual secondsSafety-cap hits
Original longer0pendingpendingpendingpendingpendingpending0
Short3–4s/turn0pendingpendingpendingpendingpendingpending0
No duration request0pendingpendingpendingpendingpendingpending0

German

TimingMatched takesCE ↑Genuine ↑Identity cosine ↑WER ↓Recognized words/minActual secondsSafety-cap hits
Original longer0pendingpendingpendingpendingpendingpending0
Short3–4s/turn0pendingpendingpendingpendingpendingpending0
No duration request0pendingpendingpendingpendingpendingpending0

French

TimingMatched takesCE ↑Genuine ↑Identity cosine ↑WER ↓Recognized words/minActual secondsSafety-cap hits
Original longer0pendingpendingpendingpendingpendingpending0
Short3–4s/turn0pendingpendingpendingpendingpendingpending0
No duration request0pendingpendingpendingpendingpendingpending0

Spanish

TimingMatched takesCE ↑Genuine ↑Identity cosine ↑WER ↓Recognized words/minActual secondsSafety-cap hits
Original longer0pendingpendingpendingpendingpendingpending0
Short3–4s/turn0pendingpendingpendingpendingpendingpending0
No duration request0pendingpendingpendingpendingpendingpending0

Dutch

TimingMatched takesCE ↑Genuine ↑Identity cosine ↑WER ↓Recognized words/minActual secondsSafety-cap hits
Original longer0pendingpendingpendingpendingpendingpending0
Short3–4s/turn0pendingpendingpendingpendingpendingpending0
No duration request0pendingpendingpendingpendingpendingpending0

Machine-readable results · All candidate input scores and selection rules

Listen side by side

All three outputs are complete, jointly generated conversations. Starting any audio pauses the other players. Input recordings are available above and inside each comparison. Complete comparisons that ended naturally appear first; capped takes remain available and are included in the statistics.

Loading comparisons…

Exactly what changed

ConditionInline timingGlobal frame request
Original longerUnchanged original durationsUnchanged original request
Short3–4 seconds per SCRIPT turn, adjusted for word countSum of the shorter turn durations
FreeAll target duration annotations removedTokens=None

A SCRIPT turn can contain more than one grammatical sentence. The original 14, 20, 40 and 56 second scenarios retain their exact original allocations; the 20 second / four-turn case was already shorter per turn than the 56 second / eight-turn case. The free condition does not quietly keep an old global length request. A 90 second execution ceiling prevents runaway sampling; it is not included in the model input and capped outputs are explicitly marked.

The model is the same final real-podcast full fine-tune of Humaneness Voice Large 1.7 used in the original five-language demo. Its final update is 9,874. The new study reuses completed original audios and their measurements. Only the short and free conditions require new inference, unless an old baseline reached its execution cap. No model training or weight change occurs here.

True logit CFG 2.5; temperature 1; top-p .95; top-k 50; repetition penalty 1; the same two seeds 7101/7102. Conditional and neutral branches share words, target timing, voice references and generated history. The conditional stop head is unchanged. Quality, WER, voice identity and turn alignment are measured using the same EARS Medium, Parakeet Redux and Nemotron models. Timing controls operate in 80 ms codec frames; this is not a millisecond-accuracy claim.

All neural work, reference scoring, selection preparation, generation and evaluation run in one Slurm allocation. The public pages are updated by a separate metadata/upload helper, requiring no additional GPU jobs.

Full protocol · Evaluation data