BB-8 Corpus Timbre & Texture Analysis (VOICE_V7)

Status: research / directional. This note compares the whole local BB-8 reference corpus of source recordings against the current VOICE_V7 output, with a focus on timbre and texture rather than the single-clip macro-shape that earlier gap analyses studied.

It does not change the voice contract. Every recommendation here is sample-affecting and would require a new voice version (V8) plus regenerated golden WAV hashes.

It supersedes the framing — not the measurements — of bb8-inquisitive-chatty-gap-analysis.md, which measured VOICE_V6 against one clip. Many of that document's recommendations (raised pause ceiling, dash/ellipsis prosody, whistle sweep, noise/breath blend, mouth stage, performance planner) shipped in VOICE_V7. This note measures what that bought, corpus-wide, and where the residual gap actually is.

Summary

VOICE_V7 is much closer to BB-8 than the V6 analysis would predict, and the headline correction is this: the gestures the earlier analysis asked for now exist and engage — but engagement is gated by punctuation and affect, not inferred from neutral text. A plain declarative phrase renders flat, over-connected, and uniformly pure-toned; the same content with sentence punctuation and exclamation/question marks engages V7's whistle, roughness, brightness-burst, pause, and staging gestures and lands inside the BB-8 corpus distribution on most timbre/texture axes.

Three things separate dootdoot from the corpus today:

1. Neutral input stays flat (the dominant practical gap). Without internal punctuation, dootdoot fuses a phrase into one long voiced island (active fraction ~0.92, median island ~900 ms, ~0 ms internal silence) with harmonicity pinned at ~0.95. BB-8's isolated library clips sit at active fraction ~0.45 with ~230 ms events. The expressive range is real but locked behind punctuation/affect, so ordinary text input does not sound BB-8-like.
2. High-register reach is still modest, and brightness is constant rather than bursty. Even on the most expressive renders, only 6–13% of active frames clear the old ~1.1 kHz ceiling, versus 22–34% for bright BB-8 clips. dootdoot keeps a steady 2–5 kHz layer present (median share ~0.058 across all renders) where BB-8 keeps that band near silent and then bursts it (max share up to ~0.81–0.97). The brightness level now matches; the brightness behavior (rare, extreme spark over a dark body) does not.
3. Frame-to-frame animation is lower on neutral input. Median spectral flux is 0.24 vs the corpus's 0.32–0.38 — the spectrum moves less. This collapses on expressive input (flux 0.46–0.60, equal to or above BB-8), so it is the same engagement-gating problem as #1 seen on a different axis.

And three things VOICE_V7 genuinely matches:

• Warble. Tremolo rate (~3.8 Hz) and depth (~0.26) land inside the BB-8 pocket (rate ~2–4 Hz, depth ~0.1–0.32). The multi-LFO warble is doing its job.
• Brightness level. Spectral centroid (~650 Hz) and 85% rolloff (~3.7 kHz) sit between the two BB-8 sub-corpora. The V6 "too bright" reading was a single-clip artifact, not a corpus-level truth.
• Staging, when punctuation is present. The raised pause ceiling and dash/ellipsis prosody work: "hello. what? wait... no!" renders as 4 islands, active fraction 0.26, and a 964 ms internal pause — squarely BB-8-like, and impossible under V6's ~240 ms ceiling.

The next step is therefore not "add more primitives" (as the V6 note concluded) but "engage the primitives V7 already has from neutral semantics, push the high register harder, and make the upper-mid layer bursty instead of constant."

Method

The corpus splits into two acoustically distinct sub-corpora, analyzed separately:

• bb8-lib (32 clips): the numbered isolated sound effects (bb8-01.mp3 … bb8-32.mp3, plus bb2-02.mp3). Mostly single short gestures.
• bb8-ctx (7 clips): the bb8-clips/ performance excerpts cut from the supercut (alarm, anxious, surprise, sad, excited, found/fixed, inquisitive-then-chatty). Continuous performed exchanges.

Decode + render + the existing band/timing report are produced by scripts/bb8-metrics (decodes every MP3 to mono 44.1 kHz 16-bit, renders a phrase corpus via xtask bb8-metrics, writes target/bb8-metrics/bb8-metrics.md). The timbre/texture block below comes from the expanded scripts/acoustic_metrics.py (driven by scripts/acoustics), run in --aggregate mode to pool clips per group and report median [p10–p90]:

scripts/bb8-metrics <reference-recordings-dir>    # decode corpus + render baseline phrase set

# 15 emotionally-varied dootdoot renders mirroring the clip contexts:
#   target/bb8-metrics/dootdoot-expanded/*.wav
uv run scripts/acoustic_metrics.py --aggregate \
  bb8-lib=target/bb8-metrics/reference-wav/*.wav \
  bb8-ctx=target/bb8-metrics/contextual-wav/*.wav \
  dootdoot=target/bb8-metrics/dootdoot-expanded/*.wav

What the expanded metrics measure

acoustic_metrics.py gained a timbre/texture block (additive; the existing directional fields and their formulas are unchanged, so prior single-clip notes still reproduce):

Metric	What it captures	Why it matters for BB-8
spectral_flatness_median	Wiener entropy: tonal (→0) vs noisy (→1)	Separates pitched tone from breath/noise texture
spectral_rolloff_hz	85% magnitude rolloff	Brightness edge, robust to a single bright bin
spectral_bandwidth_hz	Power-weighted spread around the centroid	Timbral width — thin tone vs full-bodied buzz
zero_crossing_rate	Time-domain crossings/sec	HF/noise content cross-check on centroid
spectral_flux_median	L2 change between L2-normalized adjacent spectra	Animation / liveliness — how much the timbre moves
onset_rate_hz	Flux-peak onsets per second	Articulation density (chatter vs sustained tone)
tremolo_rate_hz / tremolo_depth	Dominant 2–20 Hz AM of the RMS envelope, and its depth	Warble — the signature BB-8 amplitude wobble
crest_factor_db	Peak vs active-frame RMS	Dynamics — peaky transients vs flat sustain

These are directional tuning instruments, not acceptance constants. MP3-sourced references, formant-heavy droid audio, and simple F0/peak tracking all add noise; the value is in the large, consistent differences, not the third significant figure.

Corpus measurements

median [p10–p90] across each group. dootdoot = 15 emotionally-varied renders.

metric	bb8-lib (n=32)	bb8-ctx (n=7)	dootdoot (n=15)	Read
duration (s)	1.52 [0.84–2.6]	3.0 [1.0–3.0]	1.18 [0.88–2.1]	comparable spans
active fraction	0.45 [0.29–0.54]	0.99 [0.68–1.0]	0.92 [0.82–0.95]	dootdoot fills the file like a performance, not like an isolated SFX
active islands	2 [1–6]	4 [1–7]	1 [1–4]	dootdoot fuses into one island on neutral input
median island (ms)	232 [110–463]	389 [158–964]	911 [279–1350]	dootdoot events are far longer / over-connected
max internal gap (ms)	64 [0–447]	12 [0–511]	0 [0–91]	neutral dootdoot rarely goes truly silent mid-clip
dominant-peak range (Hz)	1010 [345–5410]	2610 [332–7010]	431 [332–534]	neutral dootdoot peak barely moves
frac > 1.1 kHz	0.04 [0–0.41]	0.18 [0–0.28]	0 [0–0]	neutral dootdoot never reaches whistle register
harmonicity median	0.80 [0.61–0.87]	0.79 [0.64–0.87]	0.95 [0.91–0.96]	dootdoot is much more purely periodic
harmonicity IQR	0.19 [0.12–0.32]	0.27 [0.19–0.32]	0.04 [0.02–0.09]	dootdoot does not swing clean↔rough on neutral input
centroid (Hz)	846 [473–1760]	619 [203–838]	647 [621–731]	brightness level is well-matched
2–5 kHz share, median	0.033 [0.003–0.16]	0.003 [0.001–0.013]	0.058 [0.036–0.079]	dootdoot keeps upper-mid constantly present
2–5 kHz share, max	0.31 [0.08–0.93]	0.81 [0.02–0.86]	0.16 [0.09–0.26]	BB-8 bursts upper-mid far harder than dootdoot
spectral flatness	~1e-4	~2e-5	~7e-5	all strongly tonal; not a discriminator (see caveat)
rolloff 85% (Hz)	4550 [2110–6490]	1670 [1290–2620]	3690 [3550–3940]	dootdoot sits between the sub-corpora
bandwidth (Hz)	777 [427–1260]	307 [212–507]	741 [643–875]	full-bodied like the library
ZCR (Hz)	2540 [1140–4740]	1550 [560–1940]	1460 [1300–1680]	dootdoot is slightly lower-noise
spectral flux	0.32 [0.23–0.45]	0.38 [0.28–0.42]	0.24 [0.21–0.30]	neutral dootdoot is less animated
onset rate (Hz)	6.5 [4.1–8.4]	9.0 [7.8–10.7]	6.1 [5.3–6.8]	fewer articulations than chatty BB-8
tremolo rate (Hz)	3.7 [2.4–7.8]	2.1 [2.1–4.8]	3.8 [3.2–4.2]	warble rate matched
tremolo depth	0.32 [0.13–0.37]	0.12 [0.09–0.19]	0.26 [0.20–0.34]	warble depth in range
crest factor (dB)	20.2 [16.9–22.5]	18.5 [17.2–20.3]	21.8 [20.2–23.1]	dynamics comparable (slightly peakier)

The two BB-8 sub-corpora are themselves very different, and worth holding as two distinct targets: the library is sparse, bright, wide-ranging single gestures (active 0.45, centroid 846 Hz, rolloff 4.5 kHz, peak range >1 kHz); the performance clips are dense, dark-bodied, with near-silent upper-mid that bursts hard (active 0.99, centroid 619 Hz, 2–5 kHz median 0.003 but max 0.81). dootdoot currently reads as a third thing: a continuously voiced, uniformly periodic, modestly-but-constantly bright tone.

The reframe: engagement is gated by punctuation and affect

The flat corpus aggregate above is dominated by neutral, lightly-punctuated phrases. When the input carries strong terminal punctuation, V7's expressive gestures fire and the timbre/texture profile moves decisively toward BB-8. Per-clip comparison of two bright BB-8 references against two punctuated dootdoot renders:

metric	bb8-09 (lib)	found-fixed (ctx)	dd what was that?!	dd wheeee so excited!
dominant-peak range (Hz)	3682	2606	2433	2885
frac > 1.1 kHz	0.22	0.34	0.06	0.13
harmonicity IQR	0.177	0.281	0.270	0.297
2–5 kHz share, max	0.970	0.860	0.899	0.941
spectral flux	0.327	0.366	0.456	0.602
onset rate (Hz)	9.11	10.33	10.67	8.28

On these inputs dootdoot's roughness swing (harmonicity IQR), upper-mid bursting, and animation are inside or above the BB-8 range. The whistle/flourish even lifts the dominant peak above the old ceiling (6–13% of frames, vs 0% for neutral input). The machinery the V6 note asked for is present and working.

Staging responds the same way. "hello. what? wait... no!" renders as:

active fraction 0.26   islands 4   median island 157 ms   max internal gap 964 ms

That is the BB-8 library's sparsity and a ~1 s conversational pause — both impossible under V6. The raised pause ceiling, dash/ellipsis prosody, and per-clause islanding all work.

Where the residual gap actually is

1. Neutral semantics do not engage the gestures. This is the dominant practical difference. Plain text ("hello there", "playing music and dancing", "I do not want to be left behind") produces one long over-connected island with no roughness swing, no high-register motion, and constant modest brightness. The performance planner localizes affect when punctuation/markers exist, but flat declarative text saturates into a single color. Most real input is flat declarative text, so most output does not sound BB-8-like.
2. High-register reach is modest and not bursty. Even the best expressive renders clear 1.1 kHz on only 6–13% of frames (BB-8: 22–34%), and the 2–5 kHz layer is always present at ~0.058 rather than near-zero with occasional ~0.8–0.97 bursts. The brightness level is right; the brightness behavior — rare, extreme spark over an otherwise dark body — is not. This matches the V6 "wrong kind of brightness" finding, now narrowed: it is a constant-vs-bursty problem plus an under-reaching whistle, not an overall level problem.
3. Roughness floor is too clean. Neutral harmonicity sits at ~0.95 (IQR ~0.04). The noise/breath blend exists but only swings on engaged gestures. Even a small constant roughening of the neutral body would close distance to BB-8's ~0.80 median.

Recommendations (directional; a future VOICE_V8)

1. Engage expressive gestures from neutral semantics, not just punctuation. Let the performance planner derive local arousal/role contrast from the semantic PCA axes and word-to-word movement so that an unpunctuated phrase still gets a probe/reply/flourish shape, internal rests, and a roughness/whistle accent. This is the single highest-leverage change and needs no new primitives — only wider, semantics-driven engagement of the curves V7 already drives.
2. Make the upper-mid layer bursty, not constant. Lower the default 2–5 kHz mix toward the BB-8 median-near-zero, and reallocate that energy into short, high-contrast bursts on accents (target max share ~0.6–0.9, as the corpus shows). Brightness should have an envelope, not a floor.
3. Push the whistle harder and more often on accents. Raise the proportion of accent frames that clear ~1.1 kHz toward the BB-8 0.2–0.3 range, and let selected flourishes sweep the dominant peak further into the 2–4 kHz region. The capability exists; it is under-deployed in frequency and frequency-of-use.
4. Add a small constant roughness to the neutral body. A subtle always-on noise/breath or filter-mismatch term would move neutral harmonicity off ~0.95 toward BB-8's ~0.80 without needing a gesture to fire.
5. Reduce neutral over-connection. Allow short (30–80 ms) inter-word rests on neutral multi-word input so active fraction can fall from ~0.92 toward the library's ~0.45, independent of punctuation.

Non-recommendations

• Do not lower overall brightness. Centroid and rolloff already match; the problem is the shape of the upper-mid energy, not its average level.
• Do not touch the warble. Tremolo rate and depth are already in the BB-8 pocket.
• Do not chase spectral flatness. Every group is strongly tonal (flatness ~1e-4 to ~2e-5); it does not discriminate here. Harmonicity median/IQR is the better roughness instrument.
• Do not add nondeterministic randomness, a speech vocoder, or sample libraries. The residual gap is engagement and shaping of existing deterministic primitives, not a missing instrument.

Caveats

• References are MP3-decoded sound-effect clips, not lab-isolated tones; absolute values carry MP3 and reverb coloration. Conclusions rest on large, consistent cross-group differences.
• tremolo_* reads the dominant 2–20 Hz AM of the frame-RMS envelope; on multi-event clips it reports the strongest single rate, not a full modulation spectrum.
• onset_rate and spectral_flux are gate- and normalization-dependent and should be read relatively between groups computed the same way, never as absolute thresholds.
• The dootdoot phrase set is a 15-phrase hand-built sample chosen to mirror the clip contexts; it is directional, not exhaustive.