@voqalize/avatar 0.4.4 → 0.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (180) hide show
  1. package/README.md +65 -154
  2. package/dist/index.d.ts +37 -26
  3. package/dist/index.js +48 -25
  4. package/dist/pin.d.ts +2 -0
  5. package/dist/pin.js +3 -0
  6. package/dist/react.d.ts +21 -6
  7. package/dist/react.js +24 -6
  8. package/package.json +8 -67
  9. package/LICENSE-APACHE-2.0 +0 -201
  10. package/LICENSE-CC-BY-4.0 +0 -396
  11. package/assets/README.md +0 -60
  12. package/assets/tanvi.glb +0 -0
  13. package/assets/tanya.glb +0 -0
  14. package/assets/tara.glb +0 -0
  15. package/assets/tess.glb +0 -0
  16. package/assets/tushar.glb +0 -0
  17. package/client/Avatar.tsx +0 -57
  18. package/client/AvatarClient.ts +0 -818
  19. package/client/createAvatar.ts +0 -177
  20. package/client/index.ts +0 -44
  21. package/client/internal.ts +0 -132
  22. package/client/playout.ts +0 -95
  23. package/client/react.ts +0 -13
  24. package/client/supports.ts +0 -27
  25. package/client/three/assets.ts +0 -26
  26. package/client/three/budgets.ts +0 -27
  27. package/client/three/character-rig.ts +0 -1746
  28. package/client/three/createCharacter.ts +0 -109
  29. package/client/three/holds.ts +0 -69
  30. package/client/three/internal.ts +0 -55
  31. package/client/three/motion-limits.json +0 -124
  32. package/client/three/sequences.ts +0 -281
  33. package/client/three/tanvi-asset.ts +0 -6
  34. package/client/three/tanvi.ts +0 -30
  35. package/client/three/tanya-asset.ts +0 -6
  36. package/client/three/tanya.ts +0 -29
  37. package/client/three/tara-asset.ts +0 -25
  38. package/client/three/tara.ts +0 -36
  39. package/client/three/tess-asset.ts +0 -6
  40. package/client/three/tess.ts +0 -30
  41. package/client/three/tushar-asset.ts +0 -6
  42. package/client/three/tushar.ts +0 -28
  43. package/client/types.ts +0 -235
  44. package/client/useAvatar.ts +0 -74
  45. package/dist/Avatar.d.ts +0 -44
  46. package/dist/Avatar.d.ts.map +0 -1
  47. package/dist/Avatar.js +0 -7
  48. package/dist/Avatar.js.map +0 -1
  49. package/dist/AvatarClient.d.ts +0 -274
  50. package/dist/AvatarClient.d.ts.map +0 -1
  51. package/dist/AvatarClient.js +0 -712
  52. package/dist/AvatarClient.js.map +0 -1
  53. package/dist/createAvatar.d.ts +0 -135
  54. package/dist/createAvatar.d.ts.map +0 -1
  55. package/dist/createAvatar.js +0 -75
  56. package/dist/createAvatar.js.map +0 -1
  57. package/dist/index.d.ts.map +0 -1
  58. package/dist/index.js.map +0 -1
  59. package/dist/internal.d.ts +0 -34
  60. package/dist/internal.d.ts.map +0 -1
  61. package/dist/internal.js +0 -56
  62. package/dist/internal.js.map +0 -1
  63. package/dist/playout.d.ts +0 -42
  64. package/dist/playout.d.ts.map +0 -1
  65. package/dist/playout.js +0 -85
  66. package/dist/playout.js.map +0 -1
  67. package/dist/react.d.ts.map +0 -1
  68. package/dist/react.js.map +0 -1
  69. package/dist/supports.d.ts +0 -19
  70. package/dist/supports.d.ts.map +0 -1
  71. package/dist/supports.js +0 -25
  72. package/dist/supports.js.map +0 -1
  73. package/dist/three/assets.d.ts +0 -10
  74. package/dist/three/assets.d.ts.map +0 -1
  75. package/dist/three/assets.js +0 -23
  76. package/dist/three/assets.js.map +0 -1
  77. package/dist/three/budgets.d.ts +0 -17
  78. package/dist/three/budgets.d.ts.map +0 -1
  79. package/dist/three/budgets.js +0 -21
  80. package/dist/three/budgets.js.map +0 -1
  81. package/dist/three/character-rig.d.ts +0 -366
  82. package/dist/three/character-rig.d.ts.map +0 -1
  83. package/dist/three/character-rig.js +0 -1644
  84. package/dist/three/character-rig.js.map +0 -1
  85. package/dist/three/createCharacter.d.ts +0 -60
  86. package/dist/three/createCharacter.d.ts.map +0 -1
  87. package/dist/three/createCharacter.js +0 -84
  88. package/dist/three/createCharacter.js.map +0 -1
  89. package/dist/three/holds.d.ts +0 -38
  90. package/dist/three/holds.d.ts.map +0 -1
  91. package/dist/three/holds.js +0 -61
  92. package/dist/three/holds.js.map +0 -1
  93. package/dist/three/internal.d.ts +0 -52
  94. package/dist/three/internal.d.ts.map +0 -1
  95. package/dist/three/internal.js +0 -49
  96. package/dist/three/internal.js.map +0 -1
  97. package/dist/three/motion-limits.json +0 -124
  98. package/dist/three/sequences.d.ts +0 -71
  99. package/dist/three/sequences.d.ts.map +0 -1
  100. package/dist/three/sequences.js +0 -262
  101. package/dist/three/sequences.js.map +0 -1
  102. package/dist/three/tanvi-asset.d.ts +0 -7
  103. package/dist/three/tanvi-asset.d.ts.map +0 -1
  104. package/dist/three/tanvi-asset.js +0 -7
  105. package/dist/three/tanvi-asset.js.map +0 -1
  106. package/dist/three/tanvi.d.ts +0 -24
  107. package/dist/three/tanvi.d.ts.map +0 -1
  108. package/dist/three/tanvi.js +0 -24
  109. package/dist/three/tanvi.js.map +0 -1
  110. package/dist/three/tanya-asset.d.ts +0 -7
  111. package/dist/three/tanya-asset.d.ts.map +0 -1
  112. package/dist/three/tanya-asset.js +0 -7
  113. package/dist/three/tanya-asset.js.map +0 -1
  114. package/dist/three/tanya.d.ts +0 -23
  115. package/dist/three/tanya.d.ts.map +0 -1
  116. package/dist/three/tanya.js +0 -23
  117. package/dist/three/tanya.js.map +0 -1
  118. package/dist/three/tara-asset.d.ts +0 -26
  119. package/dist/three/tara-asset.d.ts.map +0 -1
  120. package/dist/three/tara-asset.js +0 -26
  121. package/dist/three/tara-asset.js.map +0 -1
  122. package/dist/three/tara.d.ts +0 -30
  123. package/dist/three/tara.d.ts.map +0 -1
  124. package/dist/three/tara.js +0 -30
  125. package/dist/three/tara.js.map +0 -1
  126. package/dist/three/tess-asset.d.ts +0 -7
  127. package/dist/three/tess-asset.d.ts.map +0 -1
  128. package/dist/three/tess-asset.js +0 -7
  129. package/dist/three/tess-asset.js.map +0 -1
  130. package/dist/three/tess.d.ts +0 -24
  131. package/dist/three/tess.d.ts.map +0 -1
  132. package/dist/three/tess.js +0 -24
  133. package/dist/three/tess.js.map +0 -1
  134. package/dist/three/tushar-asset.d.ts +0 -7
  135. package/dist/three/tushar-asset.d.ts.map +0 -1
  136. package/dist/three/tushar-asset.js +0 -7
  137. package/dist/three/tushar-asset.js.map +0 -1
  138. package/dist/three/tushar.d.ts +0 -22
  139. package/dist/three/tushar.d.ts.map +0 -1
  140. package/dist/three/tushar.js +0 -22
  141. package/dist/three/tushar.js.map +0 -1
  142. package/dist/types.d.ts +0 -149
  143. package/dist/types.d.ts.map +0 -1
  144. package/dist/types.js +0 -140
  145. package/dist/types.js.map +0 -1
  146. package/dist/useAvatar.d.ts +0 -39
  147. package/dist/useAvatar.d.ts.map +0 -1
  148. package/dist/useAvatar.js +0 -35
  149. package/dist/useAvatar.js.map +0 -1
  150. package/src/avatar.d.ts +0 -436
  151. package/src/avatar.js +0 -1378
  152. package/src/behavior.d.ts +0 -42
  153. package/src/behavior.js +0 -114
  154. package/src/camera.js +0 -29
  155. package/src/clips.js +0 -155
  156. package/src/conformance.js +0 -119
  157. package/src/emotions.js +0 -58
  158. package/src/face-core.js +0 -180
  159. package/src/face-myna.d.ts +0 -7
  160. package/src/face-myna.js +0 -786
  161. package/src/face-peep-control-plane.js +0 -167
  162. package/src/face-peep.d.ts +0 -7
  163. package/src/face-peep.js +0 -925
  164. package/src/face-wren.d.ts +0 -7
  165. package/src/face-wren.js +0 -468
  166. package/src/faces.d.ts +0 -12
  167. package/src/faces.js +0 -53
  168. package/src/gaze.js +0 -504
  169. package/src/hand.js +0 -709
  170. package/src/head.js +0 -134
  171. package/src/idle.js +0 -668
  172. package/src/interjections.js +0 -747
  173. package/src/line-art.js +0 -137
  174. package/src/params.js +0 -180
  175. package/src/perform.js +0 -107
  176. package/src/prosody.js +0 -741
  177. package/src/rig.d.ts +0 -31
  178. package/src/rig.js +0 -32
  179. package/src/speech-timing.js +0 -51
  180. package/src/visemes.js +0 -410
package/src/idle.js DELETED
@@ -1,668 +0,0 @@
1
- /**
2
- * The always-on liveness layer: breathing, blinking, and a slow head drift.
3
- *
4
- * This is the cheapest layer to build and the one people notice most. A face
5
- * that holds perfectly still for two seconds looks like a crashed process, so
6
- * this layer never stops running — not even in DEGRADED.
7
- *
8
- * Output is *additive* (except blink, which takes a max against the base lid
9
- * value so a blink always closes fully regardless of the current squint).
10
- *
11
- * What the layer does is state-shaped but the layer itself is state-blind: it
12
- * renders a PROFILE (below), and the states table in avatar.js decides which
13
- * profile is in force. Blink rate alone separates listening from thinking from
14
- * busy at a glance — docs/research-biomechanics.md §5: conversation ~16/min,
15
- * cognitive work ~25/min, visual task ~9/min. It is the cheapest state
16
- * differentiation the rig has.
17
- */
18
-
19
- import { approach } from './params.js';
20
- import { HeadPose } from './head.js';
21
-
22
- /**
23
- * An idle profile. Every state carries one (sparse — missing keys mean these
24
- * defaults). Rates may jump between states; *amplitudes* glide over ~1s inside
25
- * the layer, because an oscillator whose amplitude steps is a visible pop.
26
- *
27
- * sway 0..~1.2 head-drift / brow-drift / shoulder / torso amplitude
28
- * blinkGap [min,max] seconds between blinks — ALL blinks. A blink the
29
- * face is made to do (a gaze shift, a state change, a pause in
30
- * speech) replaces the next timed one instead of adding to it,
31
- * so this is the state's rate whatever else is going on.
32
- * slowBlink 0..1 share of timed blinks drawn slow (0.34 s) — the heavy,
33
- * deliberate blink of someone thinking. A share of the budget,
34
- * not a second clock on top of it.
35
- * breathRate multiplier on the 0.23 Hz resting cycle (13.8/min)
36
- * breathAmp breath excursion scale; cognitive load = faster AND shallower
37
- * hold { every:[s,s], dur:[s,s] } — freeze the sway (breath continues).
38
- * Limited animation's static hold: deliberate stillness read as
39
- * attention, and a bitrate saving besides.
40
- * rhythm { amp, freq } — quasi-rhythmic alternating shoulder work in
41
- * burst/pause phrasing; the "busy at the keyboard" tell.
42
- * flick { amp, every:[s,s] } — a rare, tiny, fast-dying yaw wiggle:
43
- * the "no, not this one" of someone hunting through options.
44
- * Deliberately quick (a flick, not a shake — the 1.5 Hz
45
- * impatience line is about sustained nodding, and this dies in
46
- * half a second).
47
- * shift [s,s] gap between postural weight shifts, or null to sit
48
- * perfectly still. Amplitude rides on `sway`, which is how the
49
- * cognitive states get the measured sway *suppression* under
50
- * load (§6.2) without a second knob.
51
- * settle [s,s] gap between the head's own small re-positionings, or
52
- * null for a head that only moves when something moves it.
53
- * Amplitude rides on `sway` too.
54
- */
55
- export const DEFAULT_PROFILE = {
56
- sway: 1.0,
57
- // A head re-positions oftener than it used to (2026-09-21). On a character
58
- // whose held angle is budgeted this is the half of "more movement" that
59
- // actually arrives: the angle is capped by `headHold` and the *rate* is not,
60
- // so a body that re-settles every couple of seconds reads as more alive than
61
- // one that reaches further every five. Still an event every second or two at
62
- // its fastest, which is nowhere near the 1.5 Hz line.
63
- settle: [1.4, 3.4],
64
- blinkGap: [1.9, 5.4],
65
- slowBlink: 0,
66
- breathRate: 1.0,
67
- breathAmp: 1.0,
68
- hold: null,
69
- rhythm: null,
70
- flick: null,
71
- shift: [7, 16],
72
- };
73
-
74
- const rand = ([a, b]) => a + Math.random() * (b - a);
75
-
76
- // No evoked blink within this long of the last blink. A glance out and back
77
- // is two large gaze shifts a second apart and carries one blink, not two; and
78
- // a state change that lands on a gaze shift is one event to the eyelids.
79
- const BLINK_REFRACTORY = 1.5;
80
- // How long an ordinary blink takes, open to open, in seconds. Centred on the
81
- // 208 ms "short" blink Hömke et al. gave their avatar listener
82
- // (research-biomechanics.md §3.6) — the one their listeners read as nothing
83
- // but a blink. It was 0.11–0.15 s, and in a recorded call on 2026-09-24 a
84
- // photographic lid covered that in two or three frames: open, shut, open, a
85
- // pop rather than a lid travelling, which is the flicker a reviewer calls
86
- // ghosting. A line face swallowed it; a photograph shows every frame.
87
- const BLINK_DUR = [0.19, 0.23];
88
-
89
- // How far into the current gap an odds-rolled evoked blink may land, as a
90
- // fraction of that gap. A state that moves its gaze oftener than it blinks —
91
- // SEARCHING_SCREEN hops 42 times a minute against a 10/min budget, and nearly
92
- // every hop clears the ramp — would otherwise blink at its gaze's rate rather
93
- // than its own. This lets such a blink only *move* the next one onto a shift,
94
- // which is the half worth keeping (research-biomechanics.md §5.4: a blink at a
95
- // boundary is worth several placed at random), and not add one.
96
- //
97
- // 0.80 is where the trade turns, measured on tara's targets at 30 min a state:
98
- // SEARCHING_SCREEN falls from 2.12x its authored rate to 1.06x and
99
- // REVIEWING_SCREEN from 1.36x to 1.01x, while 54% of the former's blinks still
100
- // land on a hop. Tighter values buy hundredths of rate and halve that
101
- // placement — 0.90 holds only 26%.
102
- const EVOKED_EARLIEST = 0.80;
103
-
104
- /**
105
- * Postural weight shift: the one part of this layer that is not periodic, and
106
- * the part that does most of the work.
107
- *
108
- * Everything else here is an oscillator, and an oscillator cannot make a body
109
- * look alive across a thirty-second hold. Fast enough to notice and it reads
110
- * as rocking; slow enough not to and it is indistinguishable from a still
111
- * image. A headless motion map of the listening state showed the
112
- * outer edge of the torso travelling zero pixels over 24 seconds — the sway
113
- * was there in the numbers and rendered as nothing at all.
114
- *
115
- * What a seated person actually does is re-settle: every 15-40 seconds the
116
- * trunk arrives at a slightly different resting posture over a second or two
117
- * and then *stays* there. Discrete, aperiodic, and mostly holding still, which
118
- * is also why it costs almost nothing to encode — the motion is rare rather
119
- * than small.
120
- *
121
- * The counter-turn is the detail that makes it read as a body rather than a
122
- * drift: the trunk goes one way and the head yaws slightly the other, because
123
- * a person shifting their weight keeps looking at the person they are
124
- * listening to. Without it the whole figure slides sideways as one piece.
125
- */
126
- const ZERO_POSTURE = { torsoTurn: 0, torsoLean: 0, headRoll: 0, headYaw: 0, shoulderL: 0, shoulderR: 0 };
127
- const POSTURE_KEYS = Object.keys(ZERO_POSTURE);
128
-
129
- function nextPosture(prev) {
130
- // A magnitude with a floor, and a side chosen against wherever the body
131
- // already is. The obvious version — draw each channel uniformly about zero —
132
- // was written first and measured worse than no shift at all: half the draws
133
- // land near the posture already held, so half the re-settles go nowhere, and
134
- // a mechanism that is invisible half the time reads as a body that only
135
- // moves sometimes. Nobody shifts their weight by a millimetre. The 25% that
136
- // stays on the same side is what keeps it off a left-right metronome.
137
- const mag = 0.20 + Math.random() * 0.32;
138
- const away = prev.torsoTurn > 0 ? -1 : prev.torsoTurn < 0 ? 1 : (Math.random() < 0.5 ? -1 : 1);
139
- const turn = mag * (Math.random() < 0.75 ? away : -away);
140
- const leanAway = prev.torsoLean > 0 ? -1 : 1;
141
- const drop = Math.random() * 0.09 - 0.03;
142
- return {
143
- torsoTurn: turn,
144
- torsoLean: (0.03 + Math.random() * 0.05) * leanAway,
145
- // The head tips slightly against the trunk, and yaws slightly against it
146
- // too, because a person shifting their weight goes on looking at the
147
- // person they are listening to. Without the counter-turn the whole figure
148
- // slides sideways in one piece, which is a camera move, not a body.
149
- headRoll: turn * -0.28 + (Math.random() * 0.10 - 0.05),
150
- headYaw: turn * -0.13,
151
- shoulderL: drop + turn * 0.07, // weight onto one side lifts that shoulder
152
- shoulderR: drop - turn * 0.07,
153
- };
154
- }
155
-
156
- /**
157
- * The head's settle: where a person's head goes while nothing is moving it.
158
- *
159
- * It used to be a pair of slow sines per axis, and on tara that was two
160
- * failures at once. At 0.05-0.09 Hz and under a degree it was too slow to see
161
- * as motion, so a listening face read as frozen — "a dead stare, only blinks",
162
- * in the video review that found it — and what could be seen of it was a head
163
- * that never stops gliding, which is the other failure. A listener's head does
164
- * neither: it is held, and every few seconds it re-positions — one to two
165
- * degrees, over about half a second — and is held again (head.js). Pose units
166
- * at sway 1; on tara a unit of yaw is 10.7°, of pitch 17.1°, of roll 5.7°. The
167
- * first cut asked for about a degree and a recorded call measured it at 3 px
168
- * of head travel, which is the frozen listener it was meant to fix.
169
- *
170
- * None of these is a nod. The renderer does not acknowledge on its own
171
- * (CLAUDE.md), so a settle is one move, never a down-and-back — that property,
172
- * and not a small number, is what keeps pitch off the wire's vocabulary. It was
173
- * held to under a degree as well, which measured as a listening head with no
174
- * pitch in it at all (1.05° peak over thirty seconds, against yaw's 5.5°), and
175
- * a head that only ever turns is a head on a turntable. The range is biased
176
- * upward: a chin that drifts down and stays there is the downcast read, which
177
- * on a photographed face arrives long before any other.
178
- *
179
- * **Widened again on 2026-09-21, on the owner's read that the whole figure is
180
- * too restricted.** The ceiling here is not this table — it is `headHold` in
181
- * the mixer (step 6b), which softens what the layers together ask the head to
182
- * *hold* against the angle the owner measured on each character by eye
183
- * (`motion-limits.json`). So these numbers are what the drawing gets and what a
184
- * budgeted character gets is its own measured angle, which is the arrangement
185
- * this file should have been sized against all along: on tara the widest settle
186
- * yaw is 4.9° under a 6° hold budget, the widest pitch 2.6° under 5°, and roll
187
- * stays the most conservative of the three because her roll *transition* is a
188
- * recorded defect at any angle. The `dur` is untouched — a wider move over the
189
- * same half second is a faster move, not a busier one, and the frequency is
190
- * what the movement budget is about (CLAUDE.md).
191
- */
192
- const SETTLE = { yaw: [0.22, 0.46], pitch: [-0.15, 0.11], roll: [0.14, 0.38], dur: [0.45, 0.8], switchP: 0.7 };
193
- const HOME = { headYaw: 0, headPitch: 0, headRoll: 0 };
194
-
195
- /**
196
- * The moving hold: what keeps a held head alive.
197
- *
198
- * Hold-and-move fixed the floating head, and the same reviewer then found the
199
- * holds themselves uncanny — "absolute 0-velocity holds", a head locked while
200
- * only the mouth moves. Animators met this long ago: a character that stops
201
- * dead reads as a still frame, so a held pose keeps drifting by a fraction of
202
- * its size: here a few tenths of a degree, small enough that a move still
203
- * reads as the event. It runs in every state, speech included, and yields to
204
- * a static hold, which is stillness on purpose.
205
- *
206
- * **It is not a sum of sines.** The first cut was, at 0.2-0.6 Hz, and the next
207
- * review read it at once as floating underwater: a periodic sway is a loop,
208
- * and people see loops. Human postural sway is a random walk, and it stops —
209
- * a head drifts for a few seconds, then is quiet for a couple. So each axis
210
- * glides to a new random offset every 0.6-2.2 s through two smoothing stages
211
- * (an S-shaped approach, no overshoot, no rhythm), and a gate alternates
212
- * drifting stretches with near-still ones. The body's idle drift takes the
213
- * same gate, so a quiet head does not sit on a torso still rocking under it.
214
- */
215
- const LIVE = {
216
- yaw: 0.07, pitch: 0.028, roll: 0.07, breathPitch: 0.006,
217
- retarget: [0.6, 2.2], tau: 0.38,
218
- drift: [2.0, 5.0], still: [1.2, 3.5], stillGain: 0.12,
219
- };
220
-
221
- /** Ease so the shift has no corners at either end — a weight shift accelerates
222
- * and settles, and a linear ramp between two postures reads as a slide. */
223
- const smoothstep = (x) => x * x * (3 - 2 * x);
224
-
225
- export class IdleLayer {
226
- constructor() {
227
- this.t = 0;
228
- this.enabled = true;
229
- this.profile = DEFAULT_PROFILE;
230
- this._profileRef = undefined;
231
- // How much the avatar is talking, 0..1, set by the mixer. Speech does not make
232
- // the head *busier* — sway still drops — it moves the liveness down into
233
- // the shoulders and torso, which is where a speaking body actually moves.
234
- this.talk = 0;
235
- this._nextBlink = 2 + Math.random() * 3;
236
- this._blinkT = -1;
237
- this._blinkDur = BLINK_DUR[0];
238
- this._lastBlinkAt = -Infinity;
239
- // Two incommensurate frequencies per axis so the sway never visibly loops.
240
- this._ph = [Math.random() * 9, Math.random() * 9, Math.random() * 9];
241
- // The moving hold (LIVE): per axis a random target, a first stage chasing
242
- // it and the output chasing that; and the drift/still gate.
243
- this._liveTo = [0, 0, 0];
244
- this._liveA = [0, 0, 0];
245
- this._liveB = [0, 0, 0];
246
- this._liveAt = 0;
247
- this._liveOn = true;
248
- this._liveFlip = rand(LIVE.drift);
249
- this._liveGate = 1;
250
- // Glided amplitudes (profiles set the target, these chase it).
251
- this._sway = 1;
252
- this._breathAmp = 1;
253
- // Breath is a phase integrator, not sin(t*f): rate changes must bend the
254
- // cycle, not teleport it.
255
- this._breathPh = Math.random() * Math.PI * 2;
256
- // Static-hold machine. _holdAmp glides 1 -> 0 -> 1 around each hold.
257
- this._holdUntil = 0;
258
- this._nextHold = 0;
259
- this._holdAmp = 1;
260
- // Work-rhythm burst/pause machine.
261
- this._rhythmOn = false;
262
- this._rhythmFlip = 0;
263
- this._rhythmAmp = 0;
264
- // Flick machine: _flickT is time-into-flick, <0 = idle.
265
- this._flickAt = 0;
266
- this._flickT = -1;
267
- // Weight-shift machine: a posture the body is easing from, one it is
268
- // easing to, and a long wait in between.
269
- this._postFrom = ZERO_POSTURE;
270
- this._postTo = ZERO_POSTURE;
271
- this._shiftAt = 8;
272
- this._shiftT0 = -1;
273
- this._shiftDur = 2;
274
- // Settle machine: the head's held pose, and when it next moves.
275
- this._head = new HeadPose();
276
- this._settleAt = 1 + Math.random() * 2;
277
- this._settleSide = Math.random() < 0.5 ? -1 : 1;
278
- this._settled = false;
279
- // Speech phrasing: a slow gain the talking body's excursion rides on, so
280
- // it comes in waves rather than as a steady hum.
281
- this._phrase = 0;
282
- this._phraseFlip = 0;
283
- this._phraseTo = 1;
284
- // Global amplitude on everything this layer emits. The honest answer to
285
- // "keep idle motion cheap": a deployment that composites the avatar into
286
- // an encoded stream turns it down, rather than the default being frozen.
287
- this.gain = 1;
288
- }
289
-
290
- /** Adopt a (sparse) profile. Cheap to call every frame; same ref is a no-op. */
291
- setProfile(p) {
292
- if (p === this._profileRef) return;
293
- this._profileRef = p;
294
- this.profile = Object.assign({}, DEFAULT_PROFILE, p || {});
295
- // Re-arm the hold scheduler so a state that uses holds doesn't inherit a
296
- // stale far-future timestamp from one that doesn't.
297
- this._nextHold = this.t + (this.profile.hold ? rand(this.profile.hold.every) : 0);
298
- this._flickAt = this.t + (this.profile.flick ? rand(this.profile.flick.every) : 0);
299
- // A shift already under way is left alone: interrupting a weight shift
300
- // half-finished is a lurch, and states change far more often than the
301
- // body re-settles.
302
- if (this._shiftT0 < 0) {
303
- this._shiftAt = this.t + (this.profile.shift ? rand(this.profile.shift) * 0.6 : 0);
304
- }
305
- }
306
-
307
- /**
308
- * Blink now. `evoked` is a blink the face is made to do by something else —
309
- * a large gaze shift, a state change, a pause in speech — and it is refused
310
- * inside BLINK_REFRACTORY of the last one, and again before EVOKED_EARLIEST
311
- * of the way through the current gap. Every blink, evoked or not, restarts
312
- * the timer: people blink at a rate, and an event moves a blink earlier
313
- * rather than adding one. Without that, a state that glances or wanders
314
- * blinks at its timer's rate plus its gaze's, far past the rates in
315
- * docs/research-biomechanics.md §5 that the gaps are set from.
316
- *
317
- * `forced` is a beat the state asked for by name — `glance.blinkTo` and
318
- * `blinkBack`, the re-engagement the author wanted seen. It clears the budget
319
- * gate but not the refractory: a state whose authored beats overrun its own
320
- * rate is a decision to revisit in that state, not something for the eyelids
321
- * to drop silently.
322
- */
323
- blink(evoked = false, forced = false) {
324
- if (evoked) {
325
- if (this.t - this._lastBlinkAt < BLINK_REFRACTORY) return;
326
- const gap = this._nextBlink - this._lastBlinkAt;
327
- if (!forced && this.t - this._lastBlinkAt < EVOKED_EARLIEST * gap) return;
328
- }
329
- this._startBlink(rand(BLINK_DUR));
330
- }
331
-
332
- _startBlink(dur) {
333
- if (this._blinkT >= 0) return;
334
- this._blinkT = 0;
335
- this._blinkDur = dur;
336
- this._lastBlinkAt = this.t;
337
- this._nextBlink = this.t + rand(this.profile.blinkGap);
338
- }
339
-
340
- /**
341
- * A blink placed by speech, at a pause. It also restarts the timer: a speaker
342
- * who blinks at a clause boundary does not blink again a moment later
343
- * because a clock said so, and the timer is only there to fill a long run
344
- * of speech with no pauses in it.
345
- */
346
- phraseBlink() {
347
- this.blink(true);
348
- }
349
-
350
- /** A slow, deliberate blink — reads as thinking or fatigue. */
351
- slowBlink() {
352
- this._startBlink(0.34);
353
- }
354
-
355
- _blinkValue(dt) {
356
- if (this._blinkT < 0) return 0;
357
- this._blinkT += dt;
358
- const d = this._blinkDur;
359
- const p = this._blinkT / d;
360
- if (p >= 1) {
361
- this._blinkT = -1;
362
- return 0;
363
- }
364
- // Fast close (35% of the window), slower reopen. Symmetric blinks look robotic.
365
- return p < 0.35 ? p / 0.35 : 1 - (p - 0.35) / 0.65;
366
- }
367
-
368
- update(dt) {
369
- if (!this.enabled) return { add: {}, blink: 0 };
370
- this.t += dt;
371
- const t = this.t;
372
- const pr = this.profile;
373
-
374
- if (t >= this._nextBlink) {
375
- // Re-armed here as well as in _startBlink, for the frame the timer comes
376
- // due in the middle of a blink already running.
377
- this._nextBlink = t + rand(pr.blinkGap);
378
- // A share of them slow if the state asks for it. **The rest are single,
379
- // and a paired blink is not a thing this layer does** — a share of them
380
- // used to re-close the instant the lid reached open, which at this
381
- // duration is a 7-9 Hz flutter and was read by the owner on 2026-09-21 as
382
- // a dropped frame rather than as a pair. A human doublet is two blinks a
383
- // few hundred ms apart, which is what the gap already produces when it
384
- // draws short; nothing in research-biomechanics.md asks for the rest.
385
- if (Math.random() < pr.slowBlink) this.slowBlink();
386
- else this.blink();
387
- }
388
- const blink = this._blinkValue(dt);
389
-
390
- // Static hold: sway freezes, breath does not — held breath reads as alarm.
391
- if (pr.hold) {
392
- if (t >= this._nextHold && t >= this._holdUntil) {
393
- this._holdUntil = t + rand(pr.hold.dur);
394
- this._nextHold = this._holdUntil + rand(pr.hold.every);
395
- }
396
- } else {
397
- this._holdUntil = 0;
398
- }
399
- this._holdAmp = approach(this._holdAmp, t < this._holdUntil ? 0 : 1, 0.18, dt);
400
-
401
- // Amplitudes glide; rates jump. (~1s ramp keeps profile changes silent.)
402
- this._sway = approach(this._sway, pr.sway, 1.0, dt);
403
- this._breathAmp = approach(this._breathAmp, pr.breathAmp, 1.0, dt);
404
- const a = this._sway * this._holdAmp * this.gain;
405
- // The weight shift takes the profile's amplitude but NOT the hold: a hold
406
- // freezes the body where it is, and multiplying a posture by a decaying
407
- // hold factor would spring it back to centre instead.
408
- const ps = this._sway * this.gain;
409
-
410
- this._breathPh += dt * 0.23 * pr.breathRate * Math.PI * 2;
411
- const breath = (Math.sin(this._breathPh) + 1) * 0.5 * this._breathAmp;
412
-
413
- // Weight shift. Never starts inside a static hold — the hold is the state
414
- // saying "this body is concentrating", and a re-settle mid-hold undoes it.
415
- let post = ZERO_POSTURE;
416
- if (pr.shift) {
417
- if (this._shiftT0 < 0 && t >= this._shiftAt && t >= this._holdUntil) {
418
- this._shiftT0 = t;
419
- this._shiftDur = 1.5 + Math.random() * 1.6;
420
- this._postFrom = this._postTo;
421
- this._postTo = nextPosture(this._postTo);
422
- }
423
- if (this._shiftT0 >= 0) {
424
- const k = (t - this._shiftT0) / this._shiftDur;
425
- if (k >= 1) {
426
- this._shiftT0 = -1;
427
- this._shiftAt = t + rand(pr.shift);
428
- post = this._postTo;
429
- } else {
430
- const e = smoothstep(k);
431
- post = {};
432
- for (const c of POSTURE_KEYS) {
433
- post[c] = this._postFrom[c] + (this._postTo[c] - this._postFrom[c]) * e;
434
- }
435
- }
436
- } else post = this._postTo;
437
- } else {
438
- // A state that shifts nothing still has to come home from wherever the
439
- // last one left the body, or the posture sticks across the transition.
440
- this._postTo = this._postFrom = ZERO_POSTURE;
441
- }
442
-
443
- // Speech phrasing: the talking body's excursion swells and subsides over
444
- // 0.6-1.6 s rather than humming at a constant level. Speech is phrased and
445
- // a body pushing it is phrased with it; a steady oscillation while talking
446
- // is the single most robotic thing this layer could do.
447
- if (t >= this._phraseFlip) {
448
- this._phraseFlip = t + 0.6 + Math.random() * 1.0;
449
- this._phraseTo = 0.45 + Math.random() * 0.85;
450
- }
451
- this._phrase = approach(this._phrase, this._phraseTo, 0.35, dt);
452
-
453
- const s = (i, f) => Math.sin(t * f * Math.PI * 2 + this._ph[i]);
454
-
455
- // Work rhythm: bursts of alternating-shoulder movement with pauses between,
456
- // because continuous oscillation reads as rocking and phrased oscillation
457
- // reads as *doing something*. Authored ~3x the intended excursion: at
458
- // ~2.2 Hz the shoulders' 0.19s tau attenuates the target to ~0.36 of it.
459
- let workL = 0, workR = 0, workPitch = 0;
460
- if (pr.rhythm) {
461
- if (t >= this._rhythmFlip) {
462
- this._rhythmOn = !this._rhythmOn;
463
- this._rhythmFlip = t + (this._rhythmOn ? 0.5 + Math.random() * 0.7 : 0.35 + Math.random() * 0.55);
464
- }
465
- this._rhythmAmp = approach(this._rhythmAmp, this._rhythmOn ? pr.rhythm.amp : 0, 0.15, dt);
466
- const w = Math.sin(t * pr.rhythm.freq * Math.PI * 2) * this._rhythmAmp;
467
- workL = w;
468
- workR = -w * 0.85;
469
- // A trace of the same activity in the head — eyes tracking the work.
470
- workPitch = Math.sin(t * pr.rhythm.freq * 1.6 * Math.PI * 2) * this._rhythmAmp * 0.16;
471
- } else {
472
- this._rhythmAmp = 0;
473
- }
474
-
475
- // The "not this one" flick: two fast wiggles dying exponentially. Authored
476
- // ~2.5x the intended excursion — at ~2.3 Hz the head's 0.16s tau renders
477
- // roughly 0.4 of the target.
478
- let flickYaw = 0;
479
- if (pr.flick) {
480
- if (this._flickT < 0 && t >= this._flickAt) this._flickT = 0;
481
- if (this._flickT >= 0) {
482
- this._flickT += dt;
483
- const ft = this._flickT;
484
- if (ft > 0.6) {
485
- this._flickT = -1;
486
- this._flickAt = t + rand(pr.flick.every);
487
- } else {
488
- flickYaw = pr.flick.amp * Math.sin(ft * 2.3 * Math.PI * 2) * Math.exp(-ft / 0.18);
489
- }
490
- }
491
- } else {
492
- this._flickT = -1;
493
- }
494
-
495
- // The settle. Not while talking — speech holds and moves the head itself
496
- // (prosody.js), and two layers each re-positioning it is a head with two
497
- // minds — and never inside a static hold, which is stillness on purpose.
498
- if (pr.settle && this.talk < 0.5) {
499
- if (t >= this._settleAt && t >= this._holdUntil) {
500
- this._settleAt = t + rand(pr.settle);
501
- if (Math.random() < SETTLE.switchP) this._settleSide = -this._settleSide;
502
- const m = this._sway * this.gain;
503
- this._head.moveTo({
504
- headYaw: this._settleSide * rand(SETTLE.yaw) * m,
505
- headPitch: rand(SETTLE.pitch) * m,
506
- headRoll: (Math.random() < 0.5 ? -1 : 1) * rand(SETTLE.roll) * m,
507
- }, t * 1000, rand(SETTLE.dur) * 1000);
508
- this._settled = true;
509
- }
510
- } else if (this._settled) {
511
- this._settled = false;
512
- this._head.moveTo(HOME, t * 1000, 600);
513
- }
514
- const hs = this._head.sample(t * 1000);
515
- // The moving hold (LIVE). Only half of it rides on `sway`: a state that
516
- // turns its drift down still has a living head.
517
- if (t >= this._liveAt) {
518
- this._liveAt = t + rand(LIVE.retarget);
519
- for (let i = 0; i < 3; i++) this._liveTo[i] = Math.random() * 2 - 1;
520
- }
521
- if (t >= this._liveFlip) {
522
- this._liveOn = !this._liveOn;
523
- this._liveFlip = t + rand(this._liveOn ? LIVE.drift : LIVE.still);
524
- }
525
- this._liveGate = approach(this._liveGate, this._liveOn ? 1 : LIVE.stillGain, 0.5, dt);
526
- for (let i = 0; i < 3; i++) {
527
- this._liveA[i] = approach(this._liveA[i], this._liveTo[i], LIVE.tau, dt);
528
- this._liveB[i] = approach(this._liveB[i], this._liveA[i], LIVE.tau, dt);
529
- }
530
- const live = this.gain * this._holdAmp * (0.5 + 0.5 * this._sway) * this._liveGate;
531
- const [ly, lp, lr] = this._liveB;
532
- hs.headYaw += ly * LIVE.yaw * live;
533
- hs.headPitch += (lp * LIVE.pitch - Math.sin(this._breathPh) * LIVE.breathPitch * this._breathAmp) * live;
534
- hs.headRoll += lr * LIVE.roll * live;
535
- // The body's idle drift, quieted with the head (the speech term is not).
536
- const ad = a * (0.35 + 0.65 * this._liveGate);
537
-
538
- // Speech moves the body more, and in waves. `talk` says whether sound is
539
- // being produced, `_phrase` says how hard this stretch of it is being
540
- // pushed.
541
- const say = this.talk * this._phrase;
542
-
543
- return {
544
- blink,
545
- add: {
546
- // Sway frequencies run ~1.6x the original set and amplitudes ~2x.
547
- // docs/research-biomechanics.md §6.2 recorded the old numbers as
548
- // deliberately about a quarter of measured human sway, traded for
549
- // encoder cost; at that setting the head's whole idle excursion
550
- // rendered as half a pixel at call-tile size, which is not a quiet
551
- // motion but no motion. This lands nearer half speed — still well
552
- // below the 0.04-0.6 Hz seated trunk band, still calm, but now
553
- // actually present on screen. `gain` is where a deployment that
554
- // really is paying for the pixels turns it back down.
555
- // The head is the exception: it settles rather than sways (SETTLE).
556
- // The *idle* terms below went up again on 2026-09-21 — the owner read
557
- // the figure as too restricted — and only the idle terms: the speech
558
- // share is `prosody.js`'s business and is raised there, beside the
559
- // words it is timed to. These are all far below 1.5 Hz, so the raise
560
- // buys excursion and not rate; what it costs the encoder is a slightly
561
- // larger slow motion, which is the cheap kind.
562
- headYaw: hs.headYaw + flickYaw + post.headYaw * ps,
563
- headPitch: hs.headPitch + workPitch,
564
- headRoll: hs.headRoll + post.headRoll * ps,
565
- // Speech reorganises breathing: a quick inbreath at a pause and a long
566
- // outbreath over the phrase (research-biomechanics.md §6.1), which the
567
- // prosody layer draws. Quiet breathing steps mostly back under it.
568
- breath: breath * this.gain * (1 - 0.6 * this.talk),
569
- // The brows are never quite still either — and they were nearly still
570
- // here, at an amplitude a beat could swallow whole. Two video
571
- // reviewers named a motionless upper face as the biggest reason these
572
- // read as uncanny; `prosody.POSE.brow` answered that for a face that
573
- // is *talking*, and the avatar listens far more than it speaks
574
- // (CLAUDE.md). The two rates are further apart than they were so the
575
- // pair drifts out of phase inside one call rather than over several:
576
- // a matched pair is the drawing, not the face.
577
- browRaiseL: s(0, 0.089) * 0.032 * a,
578
- browRaiseR: s(1, 0.071) * 0.032 * a,
579
- // Shoulders, and the body's share of speech emphasis — a head that
580
- // moves on its own above a torso that never does is the head-on-a-stick
581
- // read, and it was the loudest note in the first round of stakeholder
582
- // feedback. The slow pair of frequencies is deliberate now that the
583
- // amplitude is big enough to see: the old 0.31 Hz term was inaudible at
584
- // 0.012 and would have read as a twitch at 0.04.
585
- // Note which factor each term takes. The idle drift is scaled by `a`,
586
- // so a state that asks for stillness gets it. The speech term is NOT:
587
- // states that speak lower their `sway` on purpose (SPEAKING sits at
588
- // 0.55, because the head should not wander while the mouth is the thing
589
- // being watched), and running the body's share of speech through that
590
- // same number costs a third of it — measured, same seed, 30 s of
591
- // SPEAKING: shoulder line travels 6 px gated against 9 px here, and the
592
- // torso band's mean luminance range falls 28 -> 24. Sway suppression is
593
- // a statement about drift. Speech reorganises the body; it does not
594
- // park it.
595
- // The speech term is smaller than it was, and that is a handover rather
596
- // than a reduction. `prosody.js` now puts the body's share of an accent
597
- // and of a phrase's posture on the shoulders and the lean, timed off
598
- // the words; this is a pair of sine waves timed off nothing. Two
599
- // unrelated sources at full strength read as a body that is busy
600
- // without being about anything, and the one that knows what is being
601
- // said should be the louder. What is left here is the part prosody
602
- // cannot supply: motion in the gaps *between* phrases, where there is
603
- // no accent to hang anything on.
604
- shoulderL: (s(1, 0.11) * 0.5 + s(2, 0.22) * 0.5) * (0.052 * ad + 0.045 * say * this.gain)
605
- + workL + post.shoulderL * ps,
606
- shoulderR: (s(2, 0.10) * 0.5 + s(0, 0.19) * 0.5) * (0.052 * ad + 0.045 * say * this.gain)
607
- + workR + post.shoulderR * ps,
608
- torsoLean: s(0, 0.085) * (0.036 * ad + 0.035 * say * this.gain) + post.torsoLean * ps,
609
- // The trunk's own drift, small next to the weight shift that dominates
610
- // this channel. It exists so the body is not perfectly still *between*
611
- // shifts, which would make each shift read as a discrete event.
612
- torsoTurn: (s(2, 0.047) * 0.6 + s(1, 0.031) * 0.4) * 0.075 * a
613
- + post.torsoTurn * ps,
614
- },
615
- };
616
- }
617
- }
618
-
619
- /**
620
- * The listening engine is now deliberately posture-only.
621
- *
622
- * `setUserSpeaking()` receives Pipecat's VAD verdict through the client
623
- * adapter. It may alter the sustained engagement lean, but it never chooses or
624
- * fires a clip: a nod, brow acknowledgement or any other facial claim of
625
- * understanding is an explicit backend/application `action()` decision.
626
- */
627
- export class ListeningEngine {
628
- constructor() {
629
- this.enabled = false;
630
- this.t = 0;
631
- this.engage = 0;
632
- this._hasSignal = false;
633
- this._explicit = null;
634
- this._speaking = false;
635
- this._silentAt = 0;
636
- }
637
-
638
- /** Compatibility no-ops retained for existing hosts that reached these
639
- * internals indirectly. With no autonomous clips there is nothing to reset
640
- * or cede. */
641
- cede() {}
642
- get ceded() { return false; }
643
- reset() {}
644
-
645
- /** Pipecat VAD's user-speaking truth. `null` removes the engagement signal. */
646
- setUserSpeaking(b) {
647
- if (b !== null) this._hasSignal = true;
648
- this._explicit = b === null ? null : !!b;
649
- }
650
-
651
- get speaking() { return this._explicit === true; }
652
-
653
- update(dt) {
654
- this.t += dt;
655
- const t = this.t;
656
- const speaking = this.speaking;
657
-
658
- // Engagement: quick to lean in when the user starts, slow to give it up —
659
- // relaxing the moment they pause would read as relief that they stopped.
660
- if (speaking !== this._speaking) {
661
- this._speaking = speaking;
662
- if (!speaking) this._silentAt = t;
663
- }
664
- const engaged = speaking || t - this._silentAt < 8;
665
- this.engage = approach(this.engage, engaged && this._hasSignal ? 1 : 0,
666
- speaking ? 1.5 : 6.0, dt);
667
- }
668
- }