Transcript Context Engine

1. Purpose

The transcript is not merely text. It is the semantic control surface for the production.

A useful transcript object must answer:

Who said what?
When?
To whom?
In response to what?
About which entities?
With what intent?
How certain were they?
What evidence supports it?
What context is required?
What story function can it perform?
What visual and audio treatments fit it?

2. Ingest stages

Media import
→ audio extraction
→ synchronization
→ speech detection
→ transcription
→ word timestamps
→ speaker diarization
→ speaker identity mapping
→ punctuation and casing
→ disfluency detection
→ non-speech event detection
→ visual scene/shot detection
→ transcript correction
→ semantic analysis

Never discard the raw transcript. Store:

1. raw_asr_text
2. corrected_verbatim_text
3. edited_display_text
4. translation_text
5. caption_text

These are different objects. Correcting an ASR error is not the same as rewriting a quote.

3. Segmentation ladder

The engine analyzes multiple nested scales.

A segment boundary can be proposed from:

speaker change
+ long pause
+ discourse marker
+ topic embedding shift
+ named-entity shift
+ question/answer completion
+ change in emotional state
+ change in location or camera scene
+ new goal, obstacle, claim, or example
+ music or chapter cue

No single signal is sufficient in every case.

4. Context layers

Each transcript unit can request context from twelve layers.

1. Lexical context — nearby words and idioms.
2. Syntactic context — grammatical dependencies and negation.
3. Local semantic context — preceding/following clauses.
4. Conversational context — question, answer, interruption, speaker relation.
5. Topical context — definitions, entities, and subject continuity.
6. Narrative context — setup, escalation, reveal, payoff.
7. Episode context — prior chapters, promises, and conclusions.
8. Series context — recurring terminology, positions, and audience expectations.
9. World context — external facts, dates, documents, and places.
10. Audience context — assumed knowledge, objections, needs, sensitivity.
11. Visual context — what is visible but not spoken.
12. Ethical/legal context — consent, risk, privacy, defamation, and disclosure.

5. Transcript-unit schema

{
  "id": "tu_000183",
  "source_id": "cam_b_audio",
  "speaker_id": "guest_01",
  "start": 132.420,
  "end": 138.810,
  "words": [],
  "raw_text": "and then I realized it was the system not me",
  "corrected_verbatim_text": "And then I realized it was the system, not me.",
  "display_text": "I realized it was the system—not me.",
  "language": "en",
  "asr_confidence": 0.94,
  "speaker_confidence": 0.99,
  "speech_act": "realization",
  "propositions": ["prop_214"],
  "entities": ["system", "speaker_self"],
  "topics": ["responsibility", "systems"],
  "sentiment": {"valence": 0.1, "arousal": 0.55},
  "performance": {},
  "narrative_roles": ["reversal", "insight"],
  "context_dependencies": [],
  "visual_options": [],
  "risk_flags": [],
  "review_status": "reviewed"
}

6. Speech-act taxonomy

A transcript unit can have multiple acts.

• assertion
• denial
• qualification
• definition
• description
• explanation
• instruction
• command
• request
• question
• answer
• challenge
• objection
• rebuttal
• agreement
• disagreement
• confession
• disclosure
• recollection
• prediction
• warning
• promise
• recommendation
• evaluation
• joke
• punchline
• analogy
• example
• quotation
• summary
• call to action
• realization
• decision
• transition
• greeting
• sign-off

7. Claim taxonomy

8. Evidence states

UNSUPPORTED
ATTRIBUTED
SOURCE_LINKED
SOURCE_REVIEWED
CORROBORATED
DEMONSTRATED_IN_MEDIA
DISPUTED
RETRACTED
NOT_APPLICABLE

The editor must not silently convert:

"I think X" → "X"
"X may cause Y" → "X causes Y"
"one example" → "all cases"
"at that time" → "always"

9. Coreference resolution

Before extracting a standalone clip, resolve:

• pronouns: he, she, they, it, this, that;
• deictic references: here, there, then, now;
• elliptical references: "the second one," "that problem";
• shared visual references: "this chart," "what you see here";
• implicit subjects: dropped or assumed nouns;
• quoted voices: who is being imitated or quoted.

Resolution options:

expand earlier
+ include question
+ add a text bridge
+ add voice-over
+ replace pronoun in a clearly labeled paraphrase
+ reject standalone use

Do not rewrite a quotation invisibly.

10. Context-completeness test

A clip receives one point for each satisfied condition.

context_completeness
= satisfied_required_tests / total_required_tests

A standalone clip should normally meet the project threshold. High-risk claims require a stricter threshold.

11. Discourse markers

Markers help reveal logical relationships:

Cut logic must preserve the relationships these markers introduce.

12. Emotional/performance dimensions

For each unit, estimate or annotate:

valence
arousal
dominance
vulnerability
certainty
urgency
humor
anger
sadness
fear
hope
surprise
authenticity
energy
pace
pause pattern
hesitation
voice strain
laughter
overlap
gesture strength
gaze direction

These are production signals, not diagnoses of a person.

13. Visualizability dimensions

A proposition can map to:

• face/performance
• archival source
• observed action
• reenactment
• generated illustrative B-roll
• screen recording
• document highlight
• quote card
• number card
• chart
• map
• timeline
• process diagram
• comparison
• icon sequence
• kinetic typography
• abstract metaphor
• no visual replacement; hold speaker

Each option stores truth_status, cost, continuity_fit, and story_fit.

14. Open-loop graph

Questions and promises are stored as open loops.

{
  "loop_id": "loop_12",
  "opened_by": "beat_04",
  "question": "Why did the project fail despite strong demand?",
  "expected_payoff": "causal explanation",
  "resolved_by": "beat_19",
  "status": "open"
}

QA flags:

• open loop never resolved;
• payoff appears before sufficient setup;
• question repeated without escalation;
• resolution contradicts the setup;
• short-form derivative includes setup without payoff.

15. Topic segmentation logic

START new topic when:
  semantic distance exceeds threshold
  OR explicit transition marker appears
  OR goal/question changes
  OR entity cluster changes materially
  OR scene/location changes and meaning changes

DO NOT start new topic when:
  speaker changes but question remains active
  OR an example elaborates the same claim
  OR a reaction completes the same emotional beat

16. Transcript correction rules

• Preserve raw source.
• Correct names using project entities.
• Preserve dialect unless a display/caption policy calls for standardization.
• Do not remove meaningful disfluency from a legal, evidentiary, or sensitive quote.
• Store filler-word removal as an edit operation, not as transcript history deletion.
• Flag low-confidence words that alter claims, numbers, dates, names, or negation.
• Review overlapping speech manually when it changes meaning.
• Keep non-speech events that carry meaning: [laughs], [door slams], [long pause].

17. Search modes

The editor should support:

exact words
semantic meaning
speaker
topic
entity
claim type
story role
emotion/performance
question/answer pair
evidence state
visual option
risk flag
duration range
camera angle
audio quality
unused/used status

18. Output objects

The context engine emits:

• corrected transcript;
• word and utterance objects;
• topic segments;
• propositions and claims;
• speaker map;
• entity map;
• question/answer pairs;
• open loops;
• story-role candidates;
• risk flags;
• visual suggestions;
• clip boundary candidates;
• context dependencies;
• human-review queue.