One voice at a time.
Does the dialogue hold together?

The two fixed input voices

These are the exact German identity anchors the listener liked in the earlier HQ grid. English and German use the same recordings. Inputs are not model predictions.

Measured comparison

Balanced pause selection is shown below. Natural timing is the default listening view. Early results are partial; compare denominators before interpreting a mean. WER is a fraction: 0.10 means 10%.

Same conversation, different generation methods

Five scripts per language, twelve turns each, with a longer sixteen-turn fifth script. Nominal lengths are around one minute and one and a half minutes; actual generated durations are shown. Every setting is retained. Starting a player pauses the others, including other open tabs of this demo family.

What exactly is being compared?

Genuine and pleasant delivery in both models

Both checkpoints receive the same explicit instruction: a very genuine, realistic, authentic conversation that is pleasant and nice to listen to, with warmth, attentive reactions and spontaneous micro-pauses. Each turn also has the same emotion directions. GENERAL is never left blank in this pilot.

Timing

The natural-timing arm removes every inline duration and sets global Tokens=None. The timed arm uses word-count estimates of about 180 English or 168 German words per minute. These are targets, not exact duration guarantees. Execution limits of 30 seconds per utterance or 60 seconds per block are not fed to the model; every hit is retained and labelled.

Candidate selection: voice identity × content fidelity

Every block has eight seeded candidates, generated in batches of four. Best-of-1, 4 and 8 use nested prefixes of that exact pool. The single-utterance arm has one take per turn; it is not secretly reranked.

R = (0.5 × C + 0.5 × M) / (1 + WER)
C = speaker-balanced mean intended-reference identity cosine, clipped to [0,1].
M = speaker-balanced mean positive own-minus-other reference cosine,
    divided by max(1 - reference_cross_cosine, 0.05), clipped to [0,1].
Missing or unscorable turns contribute zero. No S1/S2 label swapping.

Parakeet Redux locates literal scripted words with global word timestamps. An aligned turn requires at least half its words and two words matched; embeddings need at least 1.2 seconds. C compares each turn with its intended voice, while M penalizes a single voice used for both hosts. The bounded inverse-WER factor is 1/(1+WER), avoiding a singularity at zero WER. This is candidate selection, not reinforcement training.

Pause search on the actual speech clock

For every assembled condition we try 100, 200, 300, 400 and 500 ms from the last ASR word end to the next first ASR word start. We subtract retained edge padding from inserted silence, so the gap does not accidentally count padding twice. Thirty-ms guards and nearby confident nonverbal events remain; a 5-ms fade stays outside detected speech and bursts. If a gap would cut retained edge audio it is explicitly infeasible. A labelled minimum-feasible rescue up to two seconds is tried only if all five standard gaps fail. Internal pauses within a generated block are unchanged.

Every feasible full montage is scored with EARS Medium: Genuineness-only, Audiobox Content Enjoyment-only, and a 50/50 mean after fixed training-distribution Gaussian-CDF normalization. All three winners are saved, along with fixed 400 ms. These automatic selection scores may favor peculiar pauses: listen to the fixed comparator as well.

How to read the measurements

CE is the actual Audiobox head audiobox:CE, PQ is audiobox:PQ, and Genuineness is genuineness_0_6. A historical EIV content-enjoyment proxy is retained separately in raw records and is never substituted for Audiobox CE here. Higher identity cosine indicates a closer embedding match; it is not a percent-correct measure. Correct-reference preference is the fraction of intended turns closer to their own anchor than the other anchor; missing turns count as failures. Early–late cosine compares the same intended voice across time. Nemotron-3-Diarization estimates the number of active voices.

Whole-dialogue ASR, identity and diarization are run for the normalized balanced-gap winner and fixed 400 ms when feasible. Other pause variants have whole-wave CE and Genuineness but may have no WER or speaker measurement; a dash means unmeasured, never a copied score. The table gives per-metric denominators in summary.json and includes all cases without hiding caps or failures.

Weights and sampling

The baseline is Humaneness Voice Large 1.7, the merged supervised acting release. The second model is its complete real-podcast full fine-tune at update 9,874: two epochs, 80% podcast / 20% original-format acting by presentations. This experiment does not add another adapter or change either model.

Both use true logit CFG 2.5, temperature 1, top-p .95, top-k 50 and repetition penalty 1. Conditional and neutral branches retain identical words, timing, references and acoustic history; only delivery directions differ. Seeds match across models and timing arms for corresponding units; the segment and block methods use distinct documented seed schedules because their generation units differ.

Machine-readable protocol · Measurements and denominators · Previous five-language timing study · Large 1.7 prompting manual