The GENERAL / SCRIPT format

Read the notation

An illustrative example

This example explains the grammar; it is not one of the evaluated model outputs.

GENERAL: One close-miked speaker speaks quietly and hesitantly, becoming warmer toward the end. A short breath and a light laugh are audible; the background is quiet.
SCRIPT:
(quiet, hesitant; 1.20 seconds duration) I… wasn't sure.
[0.30 seconds pause]
("A short, unvoiced inhalation through the mouth.", 0.40 seconds)
(warmer, relieved; 2.30 seconds duration) ("A soft, breathy laugh.", 0.50 seconds) But I'm glad you came.

The total elapsed budget is 4.20 s: 1.20 + 0.30 + 0.40 + 2.30. The inline laugh is already contained in the final segment. Adding it again would invent time.

Exact endpoints and overlap live in JSON

The text is readable linear notation. Real speech and vocal events can overlap. Events are inserted at the nearest word gap to their midpoint; this is an approximate prose anchor, not sample-accurate placement. Use the JSON start/end values for localization or training.

{
  "schema_version": "humanness-general-script-v1",
  "segments": [{"id": "segment-002", "start_s": 1.90, "end_s": 4.20,
                "duration_includes_inline_events": true}],
  "burst_events": [{"id": "HVA_EVENT_0001", "start_s": 1.90, "end_s": 2.40,
                    "duration_s": 0.50, "caption": "A soft, breathy laugh."}]
}

Each burst has both a start and an end. Full original captions, unmodified raw scores, reference z-scores, word endpoints, source-sentence indices, overlaps, selected settings and model revisions remain in the canonical sidecar.

Sentences, segments and pauses

Parakeet recognizes words with timestamps. The procedural planner aligns sentences to those words, then splits speech around gaps of at least 0.20 s and timeline boundaries. The demo has 443 source sentences and 662 timed speech segments. Split pieces share their source-sentence Humaneness Ears prediction, computed with approximately 0.10 s of extra context.

A segment may therefore contain only part of a sentence. Its displayed duration is a timeline budget, which may absorb subthreshold gaps or endpoint adjustments. Square-bracket pauses are inferred alignment gaps; they do not guarantee acoustic silence.

Machine-readable grammar and checks

Speech lines start with (cue; N seconds duration). Event prose is a JSON-quoted string: quotation marks, parentheses, backslashes and newlines inside model-generated prose are escaped safely. A standalone event occupies its own line. A pause occupies its own line. Only actual recognized speech sits outside the annotations.

The independent verifier parses the final text, checks positive durations, preserves every recognized word once and in order, retains every selected event once, and checks the printed elapsed budget against the audio within 0.015 s. Overlapping standalone events use the planner's shared budget representation. This proves formatting consistency; it does not prove that the words, labels or intervals are acoustically correct.

Version humanness-general-script-v1 uses combined round-bracket speech cues/durations. It extends the upstream procedural formatter, which originally emitted square-bracket speech durations. Consumers should use this versioned parser or the canonical JSON rather than assume that an older text parser accepts the new syntax.

What remains a prediction

Do not read a z-score as an emotion probability, a class softmax as calibrated confidence, or a duration as evidence that an event exists. GENERAL preserves the MOSS output verbatim, including possible unsupported claims. The reviewer notes and exact output are published together in the listening grid.