# Humaneness Voice Large 1.7: agent prompting reference

Model: https://huggingface.co/laion/Humaneness-Voice-Large-1.7
Live human manual: https://laion-humaneness-voice-large-1-7-prompting-manual.static.hf.space/

Read recipes.jsonl and filter by target, language and (for bursts) with_emotion.
Use a recipe as a validated tool configuration only when usable is true. A
provisional search winner is an experiment candidate. A tested condition can
have no qualified recipe; do not convert that failure into an endorsement.
These searches use the merged 1.7 model, no extra adapter and NO reference audio.
Do not silently substitute 1.5/1.6, a specialist LoRA, another language or decoder.

## Native input contract

Keep the literal transcript separately from the instruction. Instructions use
GENERAL followed by SCRIPT. Parenthesized text directs delivery; it must never
be added to the literal transcript. The tested round timing form is:

GENERAL: One adult speaker. Warm and tender, with gently smiling vowels.
SCRIPT:
(Let affection color the first phrase; 3.20 seconds duration) I am glad you came.
(Keep the warmth audible; 3.00 seconds duration) It means a lot to me.
(Finish softly but clearly; 2.80 seconds duration) Please stay a little longer.

The legacy square timing form is also searched and explicitly recorded. Do not
mix its evidence with the round form. For new text choose realistic sentence
durations. Frames are ceil(sum of all requested seconds / 0.08). The tested
safety cap is the requested number of frames plus 45; hitting the cap fails QA.
This is conditioning, not a guarantee of exact sentence duration. Word-aligned
duration errors are stored separately from total codec duration.

## Sampling and classifier-free guidance

Audio temperature 1.0, top_p 0.95, top_k 50, repetition penalty 1.0. CFG is applied
at each of the twelve audio channels in every 80ms frame:
guided_logits = neutral_logits + cfg * (conditional_logits - neutral_logits).
Both independent KV caches receive the SAME sampled frame. The conditional stop
head stays unchanged. CFG 1 is ordinary conditional sampling. The neutral branch
keeps the identical spoken words, language, durations and no-reference wrapper;
remove emotional/style directions. For a nonverbal burst replace its cue by a
pause of the same requested duration. Do not implement CFG by scaling LoRA weights.
Read the model repository's inference code for the native wrapper and codec.

## Reward and evidence

Emotion/style R = ((2*T + G + B)/4)/(1+WER). T is the target EARS Medium dimension,
G genuineness, B vocal-burst blending. Each uses a frozen Gaussian-CDF transform
of its own training z score; this bounded transform is not a probability of
human approval. Content Enjoyment is a diagnostic with zero reward weight.
WER is a fraction. Individual WER above 0.15 or non-natural termination makes
that take score zero. Candidate validity must be at least 0.90 and mean WER at
most 0.10. Every failed seed remains in the denominator.

Burst R uses confidence-weighted target Top-3 event time in the actual ASR-word
gap, divided by desired duration and capped at one; excess accepted burst time
is penalized, then multiply by 1/(1+WER). A separate Whisper-small locator and
commercial 768-dimensional VoiceCLAP five-head classifier measure 18 sound
classes plus no_burst. EARS's 53 event labels are a different instrument.
Weak historical detector recall is disclosed, not called production success.
No word anchor means no measurable burst success. No nominal-time fallback.

Each language/target runs three generations of nine recipes. The top three
parents produce three mutations each with a local cached language model. Six
takes score each recipe. The best wording gets a seven-point CFG scan, then
winner, runner-up and baseline get six new held-out takes each. Search,
validation and sealed texts are distinct within this experiment. Historical
English/German texts are not globally unseen project data. Sealed results do
not feed mutations. Batch shape, seed, prompt, actual outputs and failures are
retained. Batched and single-item sampling are not claimed bitwise identical.

Burst recommendations additionally need at least 0.90 measurable anchors,
two-thirds detector-supported takes, mean reward at least 0.10 and lift of
0.05 over the matched-duration pause control. These are proxy evidence gates.
Listen to the clips; detector-based fitness does not replace human judgment.

## Machine fields

Each recipes.jsonl row has schema_version, target, category, language,
with_emotion, status, usable, model, model_revision, adapter, reference, recipe,
statistics and evidence. Evidence links to all prompts, scores and MP3s, not
only cherry-picked winners. Use the named language's row. Missing recipe or
usable=false is an explicit absence of a validated recommendation.