VOICE_V2 Scope

VOICE_V2 broadens the performance contract without replacing the semantic core. The four PCA-derived knobs remain the learnable meaning layer:

1. pitch center,
2. vowel/formant position,
3. contour/glide shape,
4. warble depth.

V2 may add only deterministic, bounded performance channels around that core:

Channel	Role
Phrase timing	Boundary strength, pause length, pitch reset, declination, and pre-boundary lengthening.
Affect	Utterance valence and arousal from licensing-safe lexical and punctuation signals.
Complexity	WordPiece/character-shape scalar for articulation density and duration scaling.
Archetype	A small palette of gesture shapes such as chatter, yelp, moan, stutter/burst, and tremble.

Each channel must be a pure function of the token/control-event stream. No runtime randomness, clock input, user seed, external service, or platform-dependent data is part of the contract.

Bounds

Every V2 channel must publish fixed bounds or a finite palette. Renderers clamp to those bounds before audio generation, so every input stays inside the BB-8-family parameter space even when affect, complexity, or punctuation is extreme.

Explain Output

--explain should expose V2 channels when they help a listener learn the language. Token semantic rows remain first-class. Mood, phrase, complexity, or archetype rows may be added as control/performance rows, but they must go to stderr and must not change output routing.

Voice Rule

Any V2 channel that can alter a rendered sample belongs to the voice contract and requires regenerated golden WAV hashes. V1 assets and constants remain locked.

Implemented Phrase Prosody

The first V2 performance channel applies the pure phrase plan to synthesis:

• word boundaries keep the base syllable duration and use the existing word pause;
• clause punctuation uses 8,397 syllable samples plus the medium punctuation pause;
• sentence punctuation uses 9,371 syllable samples plus the long punctuation pause;
• phrase declination, pitch reset, final lowering, and sparse emphasis are deterministic scalar offsets applied before syllable rendering;
• consecutive punctuation keeps the first marker's glide/lengthening role while the longest single pause wins.

The VOICE_V2 CLI version string was VOICE_V2; the embedded semantic mapping is still the locked token-to-axis table now carried by assets/dootdoot_asset_v1.doot.

Implemented Affect Analysis

V2 affect analysis reads the committed MIT VADER valence table and owned arousal-signal configuration. It returns per-token valence/arousal rows plus an utterance mood:

• valence is the average normalized VADER score across matched tokens;
• arousal combines punctuation density, repeated markers, all-caps words, owned intensifiers/dampeners, token count, character/WordPiece complexity, and valence energy;
• all scores are deterministic and clamped to the documented fixed bounds.

Implemented Affect-Driven Prosody

Text analysis prepends a mood control event to the sequencer stream and adds a mood control row to --explain. Synthesis uses the utterance mood as follows:

• arousal shortens or lengthens phrase-planned syllable durations within a fixed rate window;
• arousal raises pitch register, adds warble, increases upper-mid brightness, and speeds sub-gesture motion;
• valence bends contour and biases vowel/texture brighter for positive text and darker for negative text;
• mood rows are stderr-only explain output and do not change output routing.

Implemented Complexity Scalar

V2 complexity analysis returns a bounded [0, 1] scalar from inputs the project already owns:

• non-whitespace character count;
• continuation WordPiece subtoken count beyond each word's first piece.

The scalar intentionally does not use Zipf/frequency, iconicity, or third-party VAD-style tables. Those can be considered only after the same explicit asset-license policy that governs affect assets admits them.

Implemented Complexity Articulation

Text analysis prepends a complexity control event to the sequencer stream. Synthesis uses that scalar to:

• lengthen voiced syllables by a bounded deterministic amount;
• increase sub-gesture count from one to four internal movements;
• increase articulation density for internal pitch, vowel, and sparkle motion.

The semantic four-knob meaning layer remains unchanged: complexity is a separate performance channel around the same token knobs.

Implemented Archetype Palette

V2 archetype planning selects one bounded gesture family per voiced syllable:

• chatter;
• yelp;
• moan;
• stutter/burst;
• tremble.

The selector is pure and deterministic. It considers utterance mood, complexity scalar, attached punctuation, and phrase position, then emits sparse non-vocal seasoning flags without using randomness or hidden variation.

Implemented Archetype Rendering

Text analysis inserts selected archetype control rows before voiced syllables. Synthesis uses those rows to choose finite texture paths:

• yelp adds a bright upward pitch bias and short upper transient;
• moan adds a lower, darker body texture;
• stutter/burst adds deterministic gated pulses;
• tremble adds bounded pitch and amplitude tremor;
• servo and noise-tail seasoning use deterministic oscillators, never randomness.

Frozen Contract

VOICE_V2 is frozen once the phrase, affect, complexity, and archetype channels above are present together. The final acceptance note is validation/voice-v2-expressiveness.md. The BB-8 contextual clip workflow is a directional aid for phrase-level listening checks; the committed golden WAV hashes remain the sample-level voice contract.

VOICE_V3 superseded V2 by smoothing connected phrase rendering. VOICE_V4 then smoothed repeated connected onsets, VOICE_V5 smoothed bridged word attacks, and VOICE_V6 is now the active CLI contract by smoothing repeated phrase pulsing. V2 remains the frozen expressiveness checkpoint; V3 through V6 keep those performance channels.