BB-8 Sound-Vocabulary Taxonomy (frame-by-frame) vs. dootdoot

Status: research / directional. This note segments every clip in the bb8-clips/ contextual corpus into discrete sound events (frame by frame), classifies each event into a droid sound type, and compares the resulting vocabulary against what dootdoot's synth actually emits when the same analyzer is pointed at dootdoot renders.

It does not change the voice contract. Recommendations are directional and would require a new VOICE_V* (sample-affecting) + regenerated golden hashes.

Where bb8-corpus-timbre-texture-analysis.md reports one aggregate vector per clip, this note works at the gesture level: it answers "what kinds of sounds does BB-8 use, and which ones can dootdoot make?" rather than "what is the average timbre?"

Method

scripts/sound_taxonomy.py (locked uv env, numpy/scipy, same as acoustic_metrics.py) decodes a clip, computes per-frame features (FRAME 2048 / HOP 256 ≈ 5.8 ms: dominant peak pitch, autocorrelation harmonicity, spectral centroid, 2–5 kHz share, spectral flux), then:

1. Gates to active regions (RMS ≥ 0.10·max).
2. Sub-segments each region into single gestures at boundaries — onset spikes (flux local maxima), pitch discontinuities (≥ 5 st between settled frames), and deep RMS valleys — because BB-8 chains gestures legato (continuous energy, abrupt pitch jumps), so RMS gating alone fuses a whole phrase into one island.
3. Classifies each gesture from its contour (rise / fall / arch / dip / flat / wide-sweep), tonality (tonal / mixed / noisy via median harmonicity), brightness, tremolo, and internal articulation count into a droid sound type.

# decode the 7 contextual clips, then:
uv run scripts/sound_taxonomy.py target/sound-taxonomy/ctx-wav/*.wav
uv run scripts/sound_taxonomy.py --trace clip.wav    # per-event detail
uv run scripts/sound_taxonomy.py --frames clip.wav   # raw per-frame dump

Caveats. Simple argmax peak-pitch tracking on MP3-decoded, formant-heavy, occasionally polyphonic film audio is noisy; the pitch track is 5-tap median-filtered to reject octave errors, but absolute pitch is directional, not exact. Three clips are quiet/low and partly ambient-dominated: lost-friends-sad (centroid ~140 Hz — a soft low moan under film bass) and explosion-surprise (a loud ~65 Hz explosion boom with BB-8's reaction layered on it) sit below BB-8's normal register, so their "events" are not all droid gestures. Read the large, consistent cross-group differences, not the third significant figure.

The BB-8 gesture vocabulary (90 events across 7 clips)

Tally of sound types found frame-by-frame in bb8-clips/:

Type	n	What it is
blip	30	Short (<150 ms) discrete tonal pip, near-flat. The connective tissue — BB-8 fires loose strings of these.
chirp-up	10	Short rising pip.
rising-swoop	9	Sustained wide upward glide (mid register).
rising-whistle	8	Wide upward glide into the bright/high register (>1.2 kHz, up to ~4.3 kHz).
falling-swoop	8	Sustained wide downward glide.
falling-whistle	7	Wide downward glide from a high peak (often a full octave-plus drop).
chirp-down	7	Short falling pip.
trill/chatter	4	Densely re-articulated run (many internal onsets).
steady-tone	4	Flat sustained note — incl. the high alarm siren tones (4.8 kHz, 6.8 kHz).
noise-burst	2	Noisy / inharmonic burst (rough squawk).
wide-warble	1	Flat-ish but spanning a wide band with strong AM.

Aggregate profile (90 events): contour rise 31% / fall 28% / flat 41%; tonality tonal 82% / mixed 16% / noisy 2%; 27% bright; 41% are wide-sweeps (≥12 st); gesture duration 87 ms median / 157 ms p90; per-gesture pitch range 7.3 st median / 44.9 st p90; dominant-peak pitch 689 Hz median / 1600 Hz p90; peak reach 818 Hz median / 2756 Hz p90.

Per-clip frame-by-frame character

• enemy-approaching-alarm (19 events): a scatter of mid blips and short swoops, resolving into two very high steady tones (4.8 kHz, 6.8 kHz) — the literal alarm siren. Wide range, bright tail.
• excited-explanation (22 events, the richest): rapid alternation of wide rising swoops and dramatic falling whistles (peaks crashing from highs down to 65–194 Hz, single-gesture ranges of 45–81 st), capped by a noise-burst. This is the "greatest variety" clip — even more than inquisitive.
• found-fixed-excitement (19 events): two sustained high held tones (1.3 kHz) — a happy "taa-daa" — surrounded by rising whistles/swoops and a closing chatter run.
• inquisitive-then-chatty (11 events): opens with falling swoops (downward "inquisitive" inflection), then climbs into rising whistles to ~4.3 kHz as it gets chatty.
• left-behind-anxious (7 events): wide swoops with mixed/noisy tonality (audible roughness) plus a noise-burst — the anxious texture.
• lost-friends-sad (4 events): low (65–258 Hz), slow, descending — a mournful downglide (ambient-influenced; see caveats).
• explosion-surprise (8 events): dominated by the low explosion boom (~65 Hz throughout); BB-8's reaction is not cleanly separable.

The signature is: mostly short discrete tonal blips strung loosely, punctuated by wide bidirectional glides — rising AND falling whistles that ride the dominant spectral peak up into the 1.5–7 kHz region — with occasional rough/noisy bursts for agitation.

dootdoot through the same lens (63 events across 10 renders)

Renders mirror the seven contexts plus neutral/staged phrases (e.g. enemy approaching, get ready!, what was that?!, hello. what? wait... no!).

Type	n		Type	n
blip	15		rising-whistle	4
trill/chatter	13		rising-swoop	3
chirp-up	9		warble-tone	2
chirp-down	5		falling-tone	1
rising-tone	5		wide-warble	1
steady-tone	4		falling-swoop	1

Aggregate profile (63 events): contour rise 40% / fall 13% / flat 48%; tonality tonal 86% / mixed 14% / noisy 0%; 22% bright; 21% wide-sweep; gesture duration 134 ms median / 180 ms p90; per-gesture pitch range 4.8 st median / 13.6 st p90; dominant-peak pitch 452 Hz median / 689 Hz p90; peak reach 581 Hz median / 711 Hz p90.

Side-by-side vocabulary profile

Axis	BB-8 (90 ev)	dootdoot (63 ev)	Read
contour rise / fall / flat	31 / 28 / 41%	40 / 13 / 48%	dootdoot is rising-biased; ~half the falling gestures
tonality tonal / mixed / noisy	82 / 16 / 2%	86 / 14 / 0%	dootdoot never crosses into noisy
bright (≥1.1 kHz or upmid)	27%	22%	level matches (confirms prior corpus note)
wide-sweep (≥12 st)	41%	21%	dootdoot makes half as many wide pitch excursions
gesture duration med / p90	87 / 157 ms	134 / 180 ms	dootdoot gestures ~50% longer — over-connected
per-gesture range med / p90	7.3 / 44.9 st	4.8 / 13.6 st	BB-8 swoops span ~4 octaves; dootdoot caps near 1
peak reach med / p90	818 / 2756 Hz	581 / 711 Hz	dootdoot's dominant peak almost never rides high

Mapping: which BB-8 sounds can dootdoot make?

BB-8 gesture	dootdoot support	Notes
Short blip	✅ yes	But dootdoot fires fewer (15 vs 30) and binds them into trill runs.
Rising chirp	✅ yes	chirp-up is dootdoot's most native gesture.
Falling chirp	⚠️ weak	Present (5) but smaller; falls come only from punctuation glides (−3 st).
Rising swoop	✅ partial	Exists; range capped (~13 st vs BB-8's 30–50 st).
Falling swoop	⚠️ weak	Only 1 observed vs BB-8's 8. dootdoot has no strong downward glide primitive.
Rising whistle	⚠️ partial	Whistle sweep exists (→3.4 kHz target) but under-engaged: dootdoot's dominant peak stays ≤711 Hz p90 because sparkle/whistle is a thin 4.5% layer, never the loudest component.
Falling whistle	❌ missing	0 observed. The whistle primitive only sweeps up; there is no high-peak-then-drop gesture.
Trill / chatter	✅ over-represented	dootdoot's legato over-connection produces more of these than BB-8.
Steady tone	✅ yes	But never the high siren tones (4.8–6.8 kHz).
Warble tone	✅ yes	Tremolo rate/depth already in the BB-8 pocket (prior note).
Noise-burst / rough squawk	❌ missing	dootdoot is harmonic/formant-based; noise/breath is a subordinate blend, never a true noisy burst. 0% noisy events.

Findings (ranked by leverage)

1. No falling whistle, and falling glides generally are scarce. BB-8 spends 28% of gestures falling (incl. 7 falling-whistles, 8 falling-swoops); dootdoot spends 13% and produced zero falling-whistles. Architecturally the whistle sweep only climbs (toward WHISTLE_TARGET_HZ 3.4 kHz) and archetype/internal pitch gestures are rise-weighted; the only downward motion is the modest −3 st punctuation glide. A downward-whistle/swoop gesture is the single clearest missing word in the vocabulary.
2. High-register reach is rare and brief, so the dominant peak stays low. dootdoot's brightness level matches (27% vs 22% bright), but its dominant spectral peak tops out at ~711 Hz (p90) while BB-8's rides to 1.6–2.7 kHz (p90) and beyond. dootdoot has two separate high-register mechanisms and neither carries the dominant peak in practice: the whistle sweep (whistle_sweep_pitch_hz) does move the oscillator fundamental up to 3.4–4.2 kHz — exactly the right mechanism — but it is gated to TerminalFlourish and rare body accents (tension ≥ 0.75) and only over the last 55% of one syllable (CURVE_WHISTLE_START_FRACTION 0.45), so few frames land high; the upper-mid sparkle is a separate additive layer capped at UPPER_MID_SPARKLE_MIX (0.045), subordinate to voiced by design and never the dominant peak (synth.rs:1336–1350). So the fix is engaging the whistle sweep more often, earlier, and harder, not boosting the sparkle. (Consistent with the corpus note's "constant not bursty" finding, now localized to gestures.)
3. Per-gesture pitch excursions are too small. BB-8 routinely swoops 30–50 st in a single gesture (p90 = 45 st); dootdoot caps near 14 st (the ±16 st WIDE_GESTURE_PITCH_SPAN is the hard ceiling and is rarely reached). BB-8 makes dramatic octave-spanning sweeps; dootdoot makes gentle bends.
4. dootdoot over-connects and over-articulates. Gestures run ~50% longer (134 vs 87 ms median) and lean to trill/chatter (13) where BB-8 leans to discrete blip (30). BB-8's idiom is short pips with air between them.
5. No rough/noisy register. BB-8 uses noisy/mixed tonality for agitation (anxious, the tail of excited); dootdoot stays 0% noisy. The deterministic noise/breath blend exists but never reaches burst level.

Recommendations (directional; a future VOICE_V*)

These reuse machinery dootdoot already has — the gap is engagement, polarity, range, and mix, not missing instruments (matching the prior note's conclusion). Each is sample-affecting and would require a version bump (VOICE_V9→V10) plus regenerated golden WAV hashes. They are ordered by leverage; (1) and (2) are the two that would most change how dootdoot reads.

1. Add a downward whistle/swoop (the clearest missing word)

Gap. BB-8: 7 falling-whistles + 8 falling-swoops across 7 clips (15 in the library); dootdoot: 0 falling-whistles, 1 falling-swoop. dootdoot's whistle only ever climbs.

Why, in code. whistle_sweep_pitch_hz (synth.rs:433) always sweeps from start_hz toward a target at or above WHISTLE_TARGET_HZ (3.4 kHz); there is no path that sweeps below start_hz. Downward pitch motion comes only from the −3 st PUNCTUATION_GLIDE_SEMITONES final glide and the modest Moan archetype offset — both far too small to read as a falling whistle. The role curves that do carry negative pitch_velocity (Hesitation −0.15, Aside −0.10, performance.rs:433/444) are exactly the roles where whistle_amount returns 0 (synth.rs:1090), so falling intent and whistle engagement never co-occur.

Concrete change. Generalize the sweep to a signed direction driven by the sign of pitch_velocity: when velocity ≥ 0 sweep up to the existing WHISTLE_TARGET_HZ; when velocity < 0 sweep down to a new WHISTLE_FLOOR_HZ (≈ 280–350 Hz, matching the falling-whistle landings of 65–194 Hz observed but kept inside dootdoot's bounded register). Then give the statement/exclamation TerminalFlourish a negative terminal pitch_velocity while the question Probe/flourish keeps positive — which is also semantically right (questions rise, statements fall) and mirrors the existing ±-glide split. Keep the neutral (tension == 0) path byte-identical so unchanged inputs don't move.

Expected effect. Adds a whole gesture class; brings the fall fraction from 13% toward BB-8's ~28%, and makes terminal statements land with the characteristic BB-8 descending whistle instead of a small drop.

2. Engage the whistle sweep more often, earlier, and harder

Gap. dootdoot's dominant peak p90 is ~711 Hz vs BB-8's ~2.7 kHz; only the rare flourish frame clears 1.1 kHz.

Why, in code. The fundamental-moving mechanism already exists and is correct — it is just under-deployed. whistle_amount (synth.rs:1074) is non-zero only for TerminalFlourish, or for ChattyReply/Probe once archetype_tension ≥ WHISTLE_ACCENT_TENSION_THRESHOLD (0.75) — a high bar that neutral and mildly-expressive text rarely clears. When it does fire, the sweep only starts at CURVE_WHISTLE_START_FRACTION (0.45) of the syllable, so at most ~55% of one syllable's frames land high. The sparkle layer is not the lever here (see finding 2); do not raise UPPER_MID_SPARKLE_MIX.

Concrete change. (a) Lower WHISTLE_ACCENT_TENSION_THRESHOLD and/or let the planner push body-syllable tension higher on semantic accents so accents in ordinary text reach the whistle band — raising the share of frames clearing 1.1 kHz toward BB-8's 0.2–0.3. (b) Start the sweep earlier (drop CURVE_WHISTLE_START_FRACTION toward ~0.25) on the most salient accent so more frames ride high. (c) Keep WHISTLE_PITCH_CEILING_HZ (4.2 kHz) — reach is fine; frequency of use and dwell time are the gap.

Expected effect. Moves the dominant-peak p90 up toward the 1.5–2.7 kHz BB-8 band without touching brightness level — turns the whistle from a rare punctuation ornament into a regular accent gesture.

3. Widen accent pitch excursions toward 2–4 octaves

Gap. BB-8's per-gesture pitch range p90 is ~45 st (single gestures spanning 3–4 octaves); dootdoot's p90 is ~14 st.

Why, in code. Even the "wide" gesture path is capped at WIDE_GESTURE_PITCH_SPAN_SEMITONES (±16 st, synth.rs:141), selected only when whistle_amount > 0 (synth.rs:1149); ordinary syllables use PITCH_SEMITONE_SPAN (±10 st). The internal swoop/arch (INTERNAL_PITCH_SWEEP_CENTS 220, INTERNAL_PITCH_ARCH_CENTS 90) adds only ~±2 st on top. So a single dootdoot gesture can't span what BB-8 routinely does.

Concrete change. On the single highest-salience accent per phrase (the semantic-accent the planner already identifies for VOICEV8), allow the gesture span to reach ~24–36 st — either by raising the wide-gesture span on that one gesture, or by letting the whistle sweep run across a _lengthened accent syllable so the start→peak interval itself widens. Leave non-accent gestures at the current spans so the voice doesn't become uniformly wild.

Expected effect. Lifts the per-gesture range p90 from ~14 st toward the 20–45 st BB-8 region on accents, giving the dramatic swoop the corpus is full of.

4. Shorten and separate neutral gestures

Gap. dootdoot gestures run ~50% longer (134 vs 87 ms median) and skew to trill/chatter (13) where BB-8 skews to discrete blip (30 of 90). BB-8's idiom is short pips with air between them.

Why, in code. This is the over-connection the prior corpus note flagged: neutral multi-word input fuses into one long voiced island (active fraction ~0.92, ~0 ms internal gaps). Word boundaries use a tonal bridge rather than silence (sequence.rs), so gestures rarely separate.

Concrete change. On neutral (un-punctuated) multi-word input, allow short (~30–80 ms) inter-word rests instead of always bridging, and bias neutral syllable duration shorter, so the median gesture falls toward ~90 ms and active fraction toward the library's ~0.45. This is recommendation 5 of the prior corpus note, here confirmed at the gesture level.

Expected effect. Shifts the type mix from trill/chatter back toward discrete blip strings — the dominant BB-8 idiom — without removing dootdoot's ability to chatter when the text warrants it.

5. Allow an occasional rough/noisy burst on high-arousal agitation

Gap. BB-8 uses noisy/mixed tonality for agitation (anxious clip, the tail of excited; 6% genuinely noisy across the library); dootdoot is 0% noisy.

Why, in code. roughness_amount (synth.rs:1094) caps the noise/breath blend well below a true burst: even the roughest role (Hesitation, 0.7) feeds mix = amount · NOISE_BREATH_MAX_MIX = 0.35, so the voiced source still dominates and harmonicity never drops into the noisy band (≤0.42). The blend can roughen but cannot break up.

Concrete change. On a high-arousal agitated accent (negative valence + high arousal — the planner's Tremble/agitation path), let roughness_amount spike briefly toward ~1.0 (mix → NOISE_BREATH_MAX_MIX) for a short window so a single gesture crosses into the noisy band, then recovers. Keep the steady-state floor where it is; this is a transient burst, not a new baseline texture.

Expected effect. Adds the rare rough squawk BB-8 uses for surprise/agitation (noise-burst), moving a small fraction of agitated gestures into mixed/noisy without coloring the whole voice.

Non-goals (unchanged from prior analysis)

• Don't lower overall brightness (level already matches) — fix its shape and polarity.
• Don't touch warble (rate/depth already in range).
• Don't add nondeterministic randomness, a vocoder, or sample libraries; every recommendation is a deterministic re-shaping of existing primitives.