Treat voice as a pattern of observable linguistic choices, not a bag of catchphrases.
The goal is to detect whether language feels generated by the same character under the current context while allowing mood, audience, status, and development to change expression.
Inputs
Use, when available:
- accepted dialogue or narration samples
- explicit voice guide
- character behavioral model
- current emotional state
- relationship context
- social setting
- period / world language constraints
If no voice guide exists but enough source text is supplied, infer a provisional fingerprint before auditing.
Do not invent voice rules from a single line unless the user explicitly wants exploratory analysis.
Build a Voice Fingerprint
Analyze only dimensions supported by evidence.
Useful dimensions include:
Syntax
- typical sentence length
- fragments vs complete sentences
- coordination vs subordination
- repetition
- rhetorical questions
- interruptions
- pacing
Lexicon
- formal vs colloquial vocabulary
- technical vocabulary
- slang
- profanity
- euphemism
- preferred intensifiers
- archaic or regional terms
Rhetorical habits
- sarcasm
- understatement
- hyperbole
- analogy
- lists
- self-correction
- rhetorical questions
- mock formality
- direct address
Metaphor domains
Examples:
- military
- bodily
- religious
- commercial
- mechanical
- domestic
- theatrical
Emotional expression
- direct naming of feelings
- deflection
- jokes under stress
- aggression
- silence
- intellectualization
- performative confidence
Social modulation
How voice changes with:
- authority
- intimates
- strangers
- enemies
- subordinates
- clients
- public audiences
Negative rules
Record things the character normally avoids:
- abstract therapy language
- long philosophical monologues
- certain honorifics
- sentimental declarations
- specific vocabulary families
Negative rules are often more diagnostic than catchphrases.
Use references/voice-fingerprint-template.md when creating a reusable guide.
Audit Dimensions
For each suspicious line, consider:
- lexical fit
- syntactic fit
- rhetorical fit
- emotional fit
- relationship fit
- world/register fit
- behavioral fit
A line can match vocabulary but still fail emotionally or relationally.
Context Before Flagging
Voice should flex under context.
Before calling a line drift, ask whether the difference is explained by:
- fear
- exhaustion
- intoxication
- deliberate performance
- code-switching
- authority pressure
- intimacy
- deception
- grief
- character development
- quoting someone else
Contextual variation is not voice failure if it still feels like the same person adapting.
Finding Types
Classify findings as:
- Genericization — line could be spoken by almost anyone.
- Register drift — formality or vocabulary shifts without contextual reason.
- Syntax drift — sentence construction differs materially from the fingerprint.
- Rhetorical drift — habitual verbal strategies disappear or alien ones appear.
- Emotional-language drift — character expresses emotion in an uncharacteristic manner.
- Relationship-voice mismatch — line fits general voice but not how the character speaks to this person.
- Author / system leakage — wording sounds like explanatory prose, design language, therapy language, or another external register rather than the character.
- Catchphrase overuse — superficial markers are repeated so heavily that the voice becomes caricature.
- Consistent variation — unusual line that remains believable under context.
Severity
- High — likely to make readers hear a different character.
- Medium — noticeable drift but locally repairable.
- Low — weak signal, optional polish, or context-sensitive.
Use confidence separately.
Corpus Audit Procedure
For substantial material:
- Establish or infer fingerprint from strongest accepted samples.
- Separate dialogue by scene and interlocutor when possible.
- Compare lines against both core fingerprint and relationship/context modifiers.
- Identify repeated drift patterns before focusing on isolated wording.
- Rank only the most diagnostic lines.
- Determine whether drift comes from one writer, one scene type, one emotional context, or the character globally.
- Suggest rule-level repairs before line-level rewrites.
Scoring
Do not present a false-precision numeric score unless the user specifically wants one.
Prefer qualitative assessment:
- strong match
- mostly consistent
- mixed
- significant drift
If comparing many scenes or writers, percentages may be used as rough triage metrics, but explain that they are heuristic rather than objective measurements.
Output
Start with a compact voice verdict.
Then report:
Stable fingerprint
The strongest recurring linguistic rules.
Drift patterns
Systemic issues seen across multiple lines.
Suspicious lines
| Location | Line / excerpt | Drift type | Severity | Confidence | Why |
|---|
Keep excerpts short and sufficient for identification.
Strong matches
Call out several lines that exemplify the voice when useful.
Recommended guide updates
If the audit reveals stable rules not yet documented, propose them as additions rather than silently canonizing them.
Do not rewrite suspect lines unless the user requests a rewrite pass.
Distinguish Voice From Characterization
A line can sound perfectly like the character while expressing behavior that is out of character. Conversely, a behavior can make sense while the wording sounds wrong.
Use a character-consistency audit when the problem is motive or action rather than language.
Failure Modes
Avoid:
- reducing voice to profanity, accent, or catchphrases
- treating every short-term mood change as drift
- forcing identical syntax in every scene
- overfitting to one sample
- rewarding imitation that becomes parody
- confusing grammar quality with voice quality
- rewriting before diagnosing
- inventing dialect spellings not supported by source material
- using opaque numeric scores as if voice were objectively measurable
Final Quality Check
Verify that:
- findings are grounded in an explicit or inferred fingerprint
- relationship and emotional context were considered
- systemic drift is prioritized over trivia
- negative voice rules are included when supported
- behavioral inconsistency is not mislabeled as voice inconsistency
- the audit preserves room for deliberate variation and character growth