BB-8 Expressiveness Gap Analysis

Status: analysis / exploration (pre-decision). This document studies the distance between what dootdoot synthesizes today and the expressive, in-context BB-8 voice we want, and lays out the technical mechanisms that could close that distance. It is a successor to bb8-sound-signature-analysis.md, which covered the per-syllable timbre gap that Phase 7 (T-45–T-54) closed. This document covers the next layer up: phrasing, word structure, emotion, and texture — the things that make BB-8 read as communicating rather than beeping.

It does not decide anything or change VOICE_V1. It is input to a future planning pass. Where a proposal would alter output samples, this document says so. Closing these gaps belongs in VOICE_V2 (§8) and, in a few places, requires revisiting normative requirements fixed in v1.

---

1. The four gaps, stated precisely

After Phase 7, dootdoot is a structurally faithful BB-8 timbre: a deterministic formant voice with portamento, compound warble, body, sparkle, and a faint electronic edge (see voice-tuning.md). The remaining shortfall is expression and structure. The user named four gaps:

#	Gap (as observed)	One-line technical restatement
A	"Everything sounds staccato and uniformly paced — no flow, conjunction, or pausing."	No phrase-level prosody: durations, pauses, and pitch baseline are uniform and context-free.
B	"All known words read as one distinct sound; simple words should sound singular, complex words compound."	Word complexity is not mapped to sonic complexity beyond the coarse WordPiece split; each syllable is one clean gesture.
C	"No sentiment — sad text should sound morose, exciting text faster / more inflected."	No affect channel: the 4 PCA axes carry lexical semantics, not emotional valence/arousal, and nothing drives prosody from sentiment.
D	"BB-8 mixes in other kinds of sounds for texture and expression."	A fixed, narrow timbre palette: one syllable archetype, no contrastive gesture types (yelps, moans, stutters, servo blips).

These gaps overlap. A, C, and D share the same root: dootdoot renders every token through one fixed syllable archetype and lays them out on a metronome. The only per-token performance variation comes from four bounded semantic knobs tuned for learnability, not performance. Section 7 proposes a single architecture that addresses all four; Sections 3–6 analyze them one at a time.

---

2. What dootdoot does today (the precise baseline)

So the gaps are measured against ground truth, here is the exact current pipeline for the structural/expressive layer (timbre internals are in bb8-sound-signature-analysis.md).

2.1 Token → knob (engine.rs, mapping.rs)

1. Tokenize (WordPiece, uncased, add_special_tokens=false); drop control tokens; recognize prosodic punctuation .!?,;: as control markers (engine.rs:98).
2. Every voiced token → a dequantized 4-axis TokenVector (mapping.rs:79).
3. One sequence baseline = weight-scaled mean of all token vectors, squashed (engine.rs:146, mapping.rs:267). This is the utterance's single "mood" center.
4. Each token's knobs = clamp(B + α·(T − B)) around that baseline (mapping.rs:291), α = [0.85, 0.90, 1.10, 1.20] for pitch/vowel/contour/warble.

The 4 knobs are the only per-voiced-token semantic quantities that drive the synth. Their ranges are [-1, 1]. Text still affects structure through token count, continuation flags, prosodic punctuation, and syllable position; the missing channel is not "text dependence" in general, but text-dependent performance beyond the semantic knobs.

2.2 Knob → syllable (synth.rs)

render_syllable_with_final_glide (synth.rs:491) renders exactly oneBASE_SYLLABLE_SAMPLES = 7497 (170 ms) buffer per token, identical structure every time: source osc → 3 formants (vowel trajectory) → body + attack transient + sparkle → ring-mod → one amplitude envelope. Internal motion (pitch swoop, vowel bloom, compound warble) exists but is a fixed micro-gesture template scaled by the knobs.

2.3 Syllable → utterance (sequence.rs)

sequence_utterance (sequence.rs:203) concatenates syllables with fixed gaps:

• WORD_PAUSE_SAMPLES = 4851 (110 ms) between words; zero between WordPiece continuation subtokens (sequence.rs:233).
• Punctuation → a fixed final glide on the prior syllable + a fixed long/medium pause.
• Fixed 30 ms lead / 90 ms trail silence.
• Portamento carries the previous token's target pitch into the next syllable's 45 ms glide (sequence.rs:218).

2.4 The contracts that bear on the four gaps

These are frozen in v1 and are what the four gaps push against:

Frozen decision	Where	Consequence for expression
Single fixed syllable duration (V1 impl.)	every token = 170 ms	uniform rhythm; no emphasis, no lengthening (Gap A)
Single fixed inter-word pause (V1 impl.)	110 ms, always	metronomic phrasing; no clause/breath structure (Gap A)
Only the 4 axes vary (NFR-16)	mapping.rs	no channel for affect or complexity (Gaps B, C)
One syllable archetype (FR-15/17)	render_syllable*	no contrastive gesture types (Gap D)
Single utterance baseline	engine.rs:146	no pitch declination, reset, or arc across a phrase (Gap A)

None of these is wrong: each buys determinism, droid identity, or learnability. Together, they define the limits the requested expressiveness now runs into. Closing the gaps is therefore a VOICE_V2 conversation that revisits a few of these choices (§8).

Spec update (resolved): the requirement that fixed the syllable duration (FR-20) has been removed, and the fixed-inter-word-pause requirement (FR-22) has been revised to allow deterministic, boundary-dependent pauses. Variable duration and pacing are therefore no longer blocked at the requirements level; the top two rows above are now VOICE_V1 implementation choices, free to change in VOICE_V2. The remaining policy questions (NFR-16, archetype palette) are unchanged. §8 reflects this.

---

3. What "real BB-8" actually does (the evidence)

Two evidence streams: how the voice was produced, and what the audio measures.

3.1 Production: emotion is authored in language, then transcoded

The production model matters for dootdoot's architecture: BB-8's emotion and timing were authored as English, then converted to chirps. During production, Ben Schwartz wrote and recorded English dialogue for BB-8's scenes; editor Lindsey Alvarez cut it to picture to "establish the timing patterns for how the droid communicated," and only then did J.J. Abrams improvise synth takes (a Bebot-style X/Y touch synth) live to that cut, fed through a talkbox performed by Bill Hader for a second layer of live vowel formants. (SlashFilm, Post Magazine, Time)

R2-D2 was built the same way: Ben Burtt wrote equivalent English lines and "performed" them through an ARP 2600 (self-oscillating filter + sample-and-hold + ring mod, pitch slides via the glide/slew processor), describing the result as ~50% human performance, 50% machine — "because there was a human performance in it, you had a sense of R2 being alive." (Hollywood Reporter, Attack Magazine) The same English-first pattern recurs in Simlish (emotional timing locked with a non-verbal pass before gibberish is layered) (ACMI).

Why this matters: dootdoot already fits this model. Text is the emotional and timing script; synthesis is the transcode. The gap is that dootdoot currently transcodes only the lexical identity of each token and discards the two things the human performers actually rendered: emotional state (Gap C) and conversational timing/phrasing (Gap A). The architecture is validated; the channels are missing.

One more production note: emotion was carried primarily by pitch contour and rhythm, not timbre — higher/rising = positive/curious, lower/falling = sad/cautious, abrupt bursts = alarm; BB-8's famous "little sad moan" on learning Poe won't return is a single short descending gesture. BB-8 is a chirp (bird-like, child-like, emotive) against R2's mechanical beep.

3.2 Acoustics: what the contextual clips measure

We analyzed the full local sample set in two layers:

• The 32 clean top-level MP3 references, using the committed scripts/bb8-metrics workflow. These reproduce the Phase 7 comparison numbers: BB-8 has longer active islands (median ≈ 290 ms vs dootdoot ≈ 186 ms), much wider dominant-peak motion (≈ 1335 Hz vs ≈ 668 Hz), and lower harmonicity (≈ 0.81 vs ≈ 0.95). Those aggregate metrics are reliable enough for directional tuning.
• The seven emotionally-labeled contextual clips from the source recordings, decoded to mono 44.1 kHz and checked with the same RMS-gated metrics plus rough autocorrelation pitch tracking. These are useful evidence for communicative context, but they are not clean lab recordings: music, SFX, and dialogue beds keep many frames "active" (the context-only run measured an active fraction near 1.0), and formant-heavy droid audio can fool simple F0 tracking. The table below is directional, not a set of golden acoustic constants.

Neutral baseline (clean refs): rough autocorrelation over the clean set clusters around 330 Hz (E4-ish), and by-ear inspection plus dominant-peak tracking show large internal motion inside many continuous bursts. Pitch-span and micro-inflection counts depend on tracker settings, but the structural point holds: a neutral BB-8 "word" is rarely a static tone; it is usually a compact compound of micro-swoops.

Per-emotion signature (directional, not exact F0 ground truth):

Emotion / clip	Directional acoustic read	One-line identity
Sad (lost-friends-sad)	low/dark, sparse, little internal animation	stripped-down low moan/blip
Excited explanation	higher, denser, broad motion, repeated chirps	fast animated babble
Found/fixed excitement	gushing density, bright relative to nearby clips	excited up-swooping chatter
Anxious (left-behind)	rougher/less harmonic, unresolved, low-to-mid	trembling unresolved complaint
Surprise (explosion)	single high yelp/stab, little phrase development	one bare gesture at extreme urgency
Alarm (enemy-approaching)	repeated sharp stabs with wide dominant motion	urgent repeated warning
Inquisitive then chatty	separated opening gesture, then denser follow-up	question-to-response phrase arc

The table points to three structural lessons:

1. Emotion occupies distinct corners of a small acoustic space whose axes are the same affective-prosody axes (§3.3): register, span, rate/density, glide speed, warble depth, brightness. These map onto dootdoot's existing knobs plus timing — which is why an affect channel is tractable (§6).
2. Pacing is itself an emotional channel. Dense, nearly continuous chatter reads as excitement; separated repeated stabs read as alarm; a pause between an inquisitive opening and a chatty follow-up reads as a phrase arc. dootdoot's single fixed 110 ms word pause cannot express any of this.
3. Compounding is normal, and de-compounding is meaningful. Single-gesture archetypes exist, but they are expressive markers such as surprise yelps or sad stripped-down moans. dootdoot has the opposite default: every token is one clean gesture regardless of complexity or affect.

3.3 The affective-prosody literature (how emotion → acoustics, generally)

The clip findings line up with the speech-emotion literature, which gives us directional parameter targets we can apply deterministically. Consolidated from Murray & Arnott (1993), Banse & Scherer (1996), and the Juslin & Laukka (2003) meta-analysis (104 studies):

Parameter	Sadness	Excitement / Joy	Anger	Fear / Anxiety	Surprise
Speech rate	slower	faster	slightly faster	much faster	faster
Mean pitch (F0)	lower	much higher	higher	very high	much higher
Pitch range	narrower	much wider	much wider	much wider	wider
Contour	falling	rising	downward, abrupt	rising	sharp rise→fall
Intensity	softer	higher	much higher	mixed	higher
Pauses	more / longer	fewer	fewer	fewer	brief then burst
Voice quality	lax, breathy	bright, tense	harsh, tense	irregular	tense
HF energy / brightness	less	more	much more	more	more
Jitter (micro-tremor)	low	low	high	high	moderate

Sources: Banse & Scherer 1996 (PDF), Juslin & Laukka 2003 (PDF), Scherer 2004 (PDF). (Murray & Arnott directions are corroborated across these; cite verbatim adjectives as "after Murray & Arnott 1993.")

The design point: mean F0, F0 range, rate, intensity, and brightness primarily index physiological arousal, not emotion identity. That is why anger/fear/joy look acoustically similar (all high-arousal) and sadness/tenderness cluster low. Valence is the harder, finer cue, carried by contour direction (falling = negative, rising = positive), voice-quality texture (tense vs breathy), and micro-regularity (jitter). Engineering guidance: render arousal with several parameters moving together; render valence with contour and texture. The local clips follow the same pattern: the excited/sad contrast is large and multi-parameter, while same-arousal alarm/surprise/anxious separate mostly on contour and texture.

---

4. Gap A — pacing, flow, and pausing (phrase-level prosody)

4.1 The mechanism gap

dootdoot has no phrase model. Every syllable is 170 ms; every word gap is 110 ms; the pitch baseline is one constant for the whole utterance. There is no declination, no pitch reset at boundaries, no pre-boundary lengthening, no emphasis, no breath groups. Natural speech — and BB-8 — is the opposite of uniform.

The TTS literature names the exact deterministic, table-drivable levers that separate "flowing" from "robotic," all of which dootdoot lacks:

• Phrase-final lengthening (Klatt's ×1.4 rule): the syllable before a boundary is lengthened ~40%. This is a primary cue distinguishing flowing from staccato speech. (Penn prosodic duration notes)
• Declination + final lowering: F0 drifts gradually down across an utterance (tens of Hz/s), with an upward pitch reset at phrase boundaries and an extra drop at the very end. Listeners expect declination; flat F0 reads as unnatural. (Pierrehumbert 1979 (PDF))
• Break-indexed pauses: pause length scales with boundary strength — minor phrase ≈ 150 ms, comma ≈ 250–500 ms, sentence ≈ 0.4–1 s. (Parlikar & Black 2012 (PDF), Festival tutorial)
• Continuation rise vs final fall: clause-internal commas take a rising boundary tone (L-H%, "more coming"); statements fall (L-L%). dootdoot has a rough version (punctuation final glide) but only per-marker, not as a phrase tune.
• Sparse prominence: accent only some tokens (nuclear stress), not all equally. Uniform per-token gestures are a robotic tell.

Game precedents reinforce the same lever: Animalese, Undertale, and Celeste all map punctuation → pause and use author-/code-controlled dramatic pauses, never a fixed inter-unit gap. (Undertale OBJ_WRITER, Celeste/FMOD)

4.2 What to build

A deterministic phrase planner that runs between tokenization and synthesis, producing per-token timing and pitch-baseline modifiers (all functions of position and punctuation, so still a pure function of the text):

1. Declination curve. Replace the single utterance baseline with a baseline that declines linearly (a frozen Hz/s slope) from phrase start, resets up at each prosodic boundary, and applies an extra final-lowering at the last syllable. This is a pitch offset layered on the semantic pitch knob — semantics still sets relative pitch, the phrase sets the global arc.
2. Variable pauses. Make the inter-word/clause pause a function of boundary strength (word vs comma vs sentence) instead of a single constant — partly already present for punctuation; extend to clause structure and a small deterministic "breath group" rule (e.g. insert a phrase boundary every N words if none occurs).
3. Pre-boundary lengthening. Allow the syllable before a pause to render longer (e.g. ×1.3). FR-20 (the fixed-duration requirement) has been removed, so this is now permitted — see §8.
4. Sparse emphasis. Mark one token per phrase (e.g. the highest-weight or highest-arousal token) for a small duration/pitch-range boost.

4.3 Contract impact

Declination, variable pauses, and emphasis are timing/pitch-offset templates — they fit the "fixed deterministic template" philosophy that Phase 7 already embraced for micro-gestures, and require only new frozen constants. Pre-boundary lengthening needs variable duration, which the removal of FR-20 now permits (revised FR-22 likewise frees the inter-word pause). All of this remains a VOICE_V2 change — the v1 golden hashes are unaffected — but it is no longer blocked at the requirements level (§8).

---

5. Gap B — word complexity → simple vs compound sounds

5.1 What we already have, and where it stops

The design intends this mapping and partly delivers it: WordPiece splits rare/long words into multiple subtokens, each rendered as a glided continuation syllable, so playing → play + ##ing is already two flowing syllables while cat is one. That is the "frequent word = one compact syllable, rare word = multi-syllable utterance" property from design §3.1.

It falls short in two concrete ways:

1. The 30k WordPiece vocab keeps many moderately-complex words whole. Common multi-syllable words (airplane, remember, because) are often a single token → a single 170 ms gesture, so they sound as "singular" as cat. Token count under-discriminates complexity.
2. Each syllable is one clean gesture. Even when a word does split, every sub-syllable is the same archetype. The reference clips often pack many pitch and formant inflections into one compact burst (§3.2). Internal richness, more than syllable count alone, is what makes a word sound "compound."

So the perception "all known words read as one distinct sound" is accurate: complexity is quantized too coarsely (token count only) and each unit is too uniform.

5.2 The mechanism to add

Compute a per-token complexity scalar from deterministic, offline-available signals and let it drive internal sub-gesture count and articulation, independent of the semantic knobs:

• Inputs (all deterministic): subword-token count (already have it), character length, and optionally rarity via the Zipf frequency scale (log10 freq-per-billion, ~1–7) from an open frequency corpus such as SUBTLEX-US (~74k words, CC-BY-SA; requires license policy before baking). Roughly complexity = f(token_count, char_len, clamp(7 − Zipf)). (Zipf scale, SUBTLEX-US)
• Effect: complexity selects how many internal sub-swoops/articulation points a syllable renders (a level-of-detail knob — the established procedural-audio pattern of "more input → more layers/segments"; Farnell, Designing Sound). A simple/common word → one clean swoop (current behavior). A rare/long word → a compound of chained sub-gestures, matching the BB-8 norm. Optionally also lengthens the syllable for complex words (couples to Gap A's variable duration).

Sound-symbolism work points in the same direction: the bouba/kiki effect reduces to two acoustic cues, spectral balance and temporal continuity (round = continuous; spiky = segmented/discontinuous). Mapping complexity to internal segmentation uses the same lever. (Ćwiek et al. 2021, Anikin 2022). Winter et al.'s ~14k-word iconicity ratings could even be baked as an extra per-token scalar if we want sound-symbolic words to render more vividly. (Winter 2024)

5.3 Contract impact

A complexity scalar is a new baked per-token value (or a runtime function of token length + a baked Zipf table) plus new synthesis logic that varies sub-gesture count. This is a VOICE_V2 change (new mapping input + new synthesis behavior) and broadens NFR-16 (more than 4 axes now vary) — but complexity is orthogonal to the learnable semantic language: it changes how articulated a word is, not its meaning-timbre, so the learnable property is preserved (§7.2). Variable sub-gesture count also softens FR-15's "single continuous formant-glide warble" — a word becomes a short micro-phrase, which is arguably what FR-15 already gestures at for multi-token words.

---

6. Gap C — sentiment / emotional expression (the affect channel)

This gap carries the most design risk because it changes the performance layer, not just the per-token sound.

6.1 Why emotion does not "come for free" from the existing semantics

A plausible shortcut would be to rely on model2vec embeddings, since sad words cluster near other sad words. That does not give us reliable emotional prosody. PCA picks the directions of maximum semantic variance over the whole 30k vocab; there is no guarantee any of the top-4 components aligns with valence or arousal. Even when one loosely correlates, it maps to an arbitrary perceptual knob (pitch/vowel/contour/warble) with no relationship to the prosodic conventions of emotion (sad → low + flat + slow + dark). The learnability-spread.md work confirms the axes carry some structured meaning, but "structured lexical meaning" ≠ "emotional prosody." Empirically, the current output has no consistent sad-sounds-sad behavior because nothing connects text sentiment to the §3.3 acoustic directions.

The production evidence says the same thing structurally: BB-8's emotion was a separate authored channel (Schwartz's English performance), not a byproduct of which words were said. dootdoot needs an explicit affect channel too.

6.2 The mechanism: a baked VAD affect vector, pooled to an utterance mood

1. Per-token affect. Bake a valence/arousal (optionally dominance) scalar per token from an offline sentiment lexicon, alongside the existing semantic vector. This is the same precompute-and-quantize pattern as the PCA table.
2. Utterance mood. Pool per-token affect (weight-scaled mean, like the semantic baseline) into a sentence-level (valence, arousal) — the "mood" of the phrase.
3. Affect → prosody, by the §3.3 directions. Drive the global performance from mood, not the per-token meaning-timbre:
   • Arousal (multi-parameter) → speech rate (syllable duration + pause length), pitch register bias, pitch range (scales how far the semantic pitch knob swings), warble depth/rate, brightness (sparkle/upper-mid mix), and sub-gesture density. High arousal = faster, higher, wider, brighter, more inflected — matching the excited/contextual clips directionally. Low arousal = slower, sparser.
   • Valence (finer) → contour-direction bias (positive → rising/up-swoop tendency; negative → falling/declination + final lowering) and voice-quality/brightness (negative → darker, lower-pass; the sad clip is the clearest local example). Strong negative + low arousal → the "morose" target: low, flat, dark, slow.

This follows the measured emotion directions (§3.2), the literature (§3.3), and the production intent (§3.1).

6.3 Lexicon choice — licensing is the deciding constraint

The richest academic VAD lexicons are not redistributable, which matters for a shipped, committed-asset binary:

Lexicon	Dimensions	Coverage	License	Shippable in a binary?
NRC-VAD v2.1	Valence, Arousal, Dominance	55k+	non-commercial, no redistribution	❌ (commercial license, no raw redist)
Warriner 2013	V, A, D	13,915	ambiguous (Springer supp.)	⚠️ needs license review
VADER	valence (−4..+4)	~7,500 (+emoji/slang)	MIT	✅ cleanest first-pass choice
AFINN	valence (−5..+5)	~3,382	ODbL	⚠️ possible, but needs database policy
SentiWordNet 3.0	pos/neg/obj	~117k synsets	CC BY-SA 4.0	⚠️ possible, but share-alike + WSD burden

Sources: NRC-VAD, VADER, AFINN, Warriner 2013.

Recommendation: make the first affect pass licensing-safe by construction. Bake VADER (MIT) for a clean valence axis and derive a coarse arousal proxy from features dootdoot already owns: punctuation density (!, repeated markers), all-caps, intensifier words from a small hand-curated MIT-compatible list, token count, and character/WordPiece complexity. That gives enough signal to render the high-level positive/negative contour and calm/agitated pacing split deterministically, with no runtime dependency and no share-alike database question.

Treat AFINN, SentiWordNet, SUBTLEX-US, and any VAD-frequency table as a second-phase asset decision after a license policy exists for committed derivative tables. Arousal is the larger expressive win (it drives rate/range/brightness together), but the strongest human-rated arousal source (NRC-VAD) cannot be redistributed raw. Options are (a) obtain a commercial license and verify whether a quantized derivative may ship, (b) use Warriner if its license clears review, or (c) keep improving the deterministic arousal proxy. The proxy is weaker than a human-rated VAD table, but it is shippable and still separates "calm" from "agitated."

6.4 Contract impact

A new baked affect table + an affect→prosody driver is a VOICE_V2 change and the biggest expansion of the input-dependent surface. It must be folded into the VOICE_V1→V2 contract (new mapping input, new synthesis behavior) and surfaced in --explain (an extra mood row supports the learnability goal). It also broadens NFR-16 — but, as with complexity, affect is a separate orthogonal channel (mood) layered over the learnable semantic gesture, so it deepens expression without dissolving the learnable language (§7).

---

7. Gap D — additional sound textures and gesture types

7.1 The palette gap

Phase 7 added body/transient/sparkle layers inside the one syllable archetype, but every token still renders the same archetype. BB-8's vocabulary is a family of related gestures: continuous warbling chatter, single high yelps (surprise), short descending moans (sadness), rapid stutters (excitement), buzzy/rough trembles (anxiety; the anxious clip measured low 0.57 harmonicity), and percussive rising stabs (alarm). The acoustic analysis points to structure, not only knob values. Surprise is one bare sustained tone, sad is one dark blip, and alarm is repeated transient-onset stabs. They are structurally different gesture archetypes.

7.2 The mechanism: a small archetype palette selected by affect/structure

Introduce a bounded set of deterministic gesture archetypes (still all within the droid parameter space), selected per token/phrase by the affect + complexity channels rather than chosen freely:

• Chatter (default) — the current continuous warbling syllable.
• Yelp — short, high, sustained, single inflection; triggered by very high arousal + surprise context (e.g. a ! after a short utterance).
• Moan — low, dark, falling glissando, slow; triggered by strong negative valence + low arousal.
• Stutter/burst — rapid sub-gesture repetition; high arousal + high complexity.
• Tremble — added jitter + slight inharmonicity (a deterministic rough texture); fear/anxiety (negative valence + high arousal).

Plus non-vocal seasoning used sparingly: short servo/mechanical blips or a brief filtered-noise breath tail at phrase boundaries — these are what give the references their "sound-effect" texture without making BB-8 read as non-vocal (R2's identity was explicitly a blend of human and machine).

This is the hardest gap to square with the project's core promise: a palette of archetypes pushes against NFR-16's "one consistent droid family" and the v1 thesis that fixedness makes the language learnable. The mitigation is to keep the palette small, bounded, and deterministically selected by the affect/complexity channels (not free variation), so the archetype itself becomes part of the learnable language ("BB-8 yelps when surprised") rather than noise. Determinism is preserved: archetype selection is a pure function of the text.

7.3 Contract impact

New synthesis archetypes + a selection rule = VOICE_V2, and the most significant reinterpretation of NFR-16 (the "bounded droid parameter space" now includes a discrete archetype dimension). Recommend doing this last and conservatively, after the affect channel exists to drive selection — an archetype palette with nothing principled selecting it would just be variety for its own sake.

---

8. Determinism and contract implications

None of the proposals threaten determinism: every new input (complexity scalar, affect vector, phrase position, archetype choice) is a pure deterministic function of the text plus frozen tables/constants, computed with the existing owned-math path. The buffer-as-source-of-truth and bit-exact guarantees (§8.1 of design.md) are untouched.

What they do require is honest contract accounting:

1. This is a VOICE_V2. Every proposal alters output samples. Per the freeze rule, the v1 golden fixtures stay as the v1 contract; v2 gets its own version id, header, and regenerated golden hashes. The VOICE_V1 lock (T-54) already established the machinery for that.
2. Requirement changes — two already resolved, one still open:
   • FR-20 (fixed syllable duration) — removed. This was the hard one: pre-boundary lengthening, complexity-driven length, and arousal-driven rate all need duration to vary, and the §3 evidence is unambiguous that uniform duration is the primary staccato cause. The requirement has been eliminated from the spec, so variable duration is now permitted. Guidance for the implementation: keep duration a function of structure and affect, never a free per-token semantic axis — so rhythm stays learnable (predictable from the text) even though it is no longer uniform.
   • FR-22 (fixed inter-word pause) — revised. Now allows deterministic, boundary-strength-dependent pauses (word vs clause vs sentence) instead of one fixed constant. Low risk; punctuation already varies pauses.
   • NFR-16 ("only the 4 bounded axes vary") — still open. Recommendation: broaden to "a fixed set of deterministic, bounded channels vary" — semantic (4 axes, learnable) + affect (valence/arousal, mood) + complexity (articulation) + archetype (gesture type). The droid-family identity is preserved by keeping every channel bounded and deterministic; it just stops being only four axes. This is the remaining requirements decision to make before the affect/complexity/archetype work.
3. New committed assets / licensing review. Start with assets that are clean to redistribute (e.g. VADER under MIT plus hand-curated heuristics). Any ODbL, CC-BY-SA, NRC, Warriner, Zipf, or VAD-derived table needs explicit license policy before it is committed and include_bytes!-embedded (§6.3). The source_manifest.toml pattern extends naturally to pin approved sources.

8.1 The central tension: learnability vs expressiveness

The v1 thesis is that fixedness makes the sound-language learnable and shareable. Every gap here adds variability. The way to reconcile that is orthogonal channels:

• Keep the 4 semantic axes as the learnable core (the "words"). Untouched.
• Add affect, complexity, and archetype as separate deterministic channels (the "performance") that shape how the words are delivered, not what they mean.

A listener still learns "this timbre = this meaning" from the semantic axes; the affect channel layers a learnable "this delivery = this mood" on top. The production path for BB-8 used the same split: a lexical/timing script (the words) performed with an emotional overlay (Schwartz's English, Abrams' and Hader's live performance). Expressiveness and learnability can coexist if expression is its own structured, deterministic layer.

The residual tradeoff noted in voice-tuning.md — "dootdoot remains cleaner and more regular than the reference set" — is what these channels would address, at the cost of a more complex but still deterministic contract.

---

9. Recommended direction and rough sequencing

Order the work by perceptual impact and dependencies. Each phase is independently shippable as part of a v2 effort and testable red-green (value tests for the planners, insta snapshots for --explain, golden-WAV hashes once frozen). This is a suggested order, not a committed plan.

1. Phrase prosody planner (Gap A) — declination + final lowering + boundary reset + variable pauses + pre-boundary lengthening. Biggest perceptual win for the least new data; no new lexicon. Already unblocked at the requirements level (FR-20 removed, FR-22 revised), so this can lead.
2. Licensing-safe affect channel (Gap C) — bake VADER valence, derive arousal from owned punctuation/case/intensifier/complexity signals, pool to mood, and drive rate/register/range/brightness/contour by the §3.3 directions. This adds emotional expression without blocking on NRC-VAD, ODbL, CC-BY-SA, or ambiguous supplemental-data licenses. Surface mood in --explain.
3. Complexity → compound articulation (Gap B) — start with WordPiece count + character length; add Zipf/frequency only after asset licensing is settled. This makes long words sound compound, is self-contained, and synergizes with arousal-driven density.
4. Gesture archetype palette (Gap D) — yelp/moan/stutter/tremble + sparing non-vocal seasoning, selected by the affect/complexity channels. Last, because it depends on the others to be principled rather than arbitrary, and is the biggest hit to the "consistent family" promise — so do it conservatively.

Cross-cutting: extend --explain to show the new channels (mood, complexity, archetype), which directly serves the learnability goal and aids by-ear tuning; and reuse the Phase 7 scripts/bb8-metrics harness, adding the per-emotion corner metrics from §3.2 as directional acceptance aids (with by-ear review still the gate).

---

10. Reliability caveats

• The seven contextual clips carry a media noise bed: event counts, active fractions, and some F0/span estimates are noise- and tracker-limited. Treat the per-emotion rows in §3.2 as directional by-ear/metric synthesis, not exact acoustic constants. The most stable local signals are relative density, obvious contour/archetype, dominant-peak motion, harmonicity, and broad spectral darkness/brightness.
• The §3.3 emotion→acoustics directions are well-corroborated across three independent reviews, but exact Murray & Arnott (1993) adjectives should be cited as "directional."
• Game-system numbers (§4.1) come from reference reimplementations / decompiled source, not vendor documentation.
• This document recommends a direction; the v1 thesis that fixedness aids learnability is real. FR-20 has been removed and FR-22 revised (§8), so variable pacing is settled; the remaining NFR-16 broadening is a tradeoff that needs an explicit decision before affect, complexity, or archetype channels become normative.

---

Appendix A — source index

Production / sound design: SlashFilm BB-8 voice · Post Magazine TFA sound editing · Time BB-8 voice · Hollywood Reporter / Attack Magazine / Designing Sound (Ben Burtt, R2-D2 ARP 2600) · ACMI (Simlish).

Affective prosody: Banse & Scherer 1996 · Juslin & Laukka 2003 · Scherer 2004 · Murray & Arnott 1993 (directional).

Synthesis prosody: Penn prosodic duration (Klatt) · Pierrehumbert 1979 (declination) · Parlikar & Black 2012 / Festival (pauses) · ToBI.

Complexity / sentiment / sound symbolism: Zipf scale · SUBTLEX-US · Ćwiek 2021 / Anikin 2022 (bouba/kiki) · Winter 2024 (iconicity) · NRC-VAD · Warriner 2013 · VADER · AFINN · SentiWordNet · Farnell, Designing Sound (procedural LOD).

Game voice systems: Animalese (Acedio / equalo) · Undertale OBJ_WRITER · Celeste (Regamey / FMOD) · Banjo-Kazooie · Ace Attorney (Capcom dev blog).

(Full URLs inline in §§3–7.)