PHONOLOGICAL FORM → GESTURE ACTIVATION
From syllable representation to articulation
Follow a selected syllable or the original four-syllable sequence through neural timing cycles into overlapping gesture ramps. Pause anywhere to inspect the same moment at each level.
Select a word
Or select a stressed syllable
Or select an unstressed syllable
Word buttons play the complete lexical syllable sequence for Tasse, Tasche, Kanne, Katze, Topf, Stuhl and Schrank. Teller (short ER) plays tE → lA; Teller (long ER) plays tE → lEA. All earlier reduced syllables and bap, gum and dip remain individually selectable. Single vowels span 105.4°–240.6° in unstressed open syllables and 105.4°–257.5° in stressed syllables. In lEA, V1 (E) spans 105.4°–223.7° and V2 (A, vocalic R) spans 206.8°–308.2°. All syllables share the same consonant rules.
01 Syllable timing
One 2D oscillator per sequence slot · 1.5 Hz
● Start · □ End · Colours match Gesture Activation. The unit circle shows the stored phase boundaries, increasing counterclockwise from 0° at the left through 90°, 180° and 270° to 360°. Phase values use at most one decimal place. Nearby markers are staggered radially at the same angle for readability. The moving dot follows the recorded oscillator; neural gesture ramps can lag behind these boundaries.
Timing rules · onset, coda, glottis, velum and syllable transitions
The dot follows the decoded (x, y) state. The ring is a phase reference, not a fitted trajectory. Each slot has its own oscillator, including the two bap occurrences.
Integrated control parameters · shared across all gestures
One value per articulatory parameter, combining the contributions of all relevant gestures. Bars show decoded positive minus negative ensemble activity, with zero at the centre (−1 to +1). These are the same differences shown in section 04, before the gains and offsets used by the articulation model in section 05.
Gesture ramp integrators · show current activations
Same moment as the oscillators above · bar scale 0–1 · exact decoded values shown above each bar.
02 Routing the gesture signals
Selection releases inhibition
Bars show recorded inhibitory control activity for one representative channel per selected syllable. Low activity releases that pathway; a gesture signal still requires the appropriate oscillator phase. Values are neural states, not binary switches.
03 Gesture activation
1D recurrent ramp integrators
Coloured dashed lines mark each gesture’s phase-window start and end in the colour of its trace. Only vowel boundaries extend across all gesture tiers; every other gesture’s boundaries stay within its own tier. Where coincident lines overlap, their colours alternate. Their approximate times are interpolated from the recorded oscillator phase; neural activation ramps can lag and decay beyond these boundaries.
Gesture timing values · time and phase
Times are estimates from the filtered oscillator recording (5 ms samples). φ is the phase-window boundary in degrees; these markers do not define an activation-amplitude threshold.
All traces share the same 0–2.5 s time axis. The cursor marks the current time; the bars show the decoded activation at that instant. These are model gesture states, not measured muscle activity.
A ramp represents the activation strength of a gesture. Its targets map that activation to one or more articulatory control parameters. Several gestures can contribute to the same parameter: the two bap lip closures, C1 and C2 within one syllable, both drive Lip closure. Conversely, one Sch gesture controls tongue-tip height, tongue-tip frontness and lip rounding together. Section 04 shows the shared parameter ensembles after this mapping.
04 Agonist–antagonist activation
Two neural ensembles per articulatory parameter
Abstract opposing muscle groups: positive and negative activities are shown separately. Their unscaled difference is shown here (−1 to +1); the articulatory model below applies parameter-specific gains and offsets. These are simulated neural representations, not measured muscle forces.
Download articulatory parameter trajectories
05 Articulatory movement
Same simulation time ·
● Tongue / lips● VelumFacing left
Model input = offset + positive group × positive gain − negative group × negative gain. Velum height: offset +0.1 (+100 trainer units), negative gain 1.3. All other offsets 0 and negative gains 1.0. Positive gains: lip closure 1.0, tongue dorsum 1.2, tongue tip 1.1, glottal aperture 1.2; all other gains 1.0. Neural group activities and timing are unchanged. Glottal opening is shown separately because its width is not visible in a midsagittal section. Scope: full closures, postalveolar Sch contour, and A, I, U, E, O, plus a velar R approximant. Sch frontness interpolates the original Trainer alveolar/postalveolar contours; this visual model does not generate frication noise. Neural timing is used unchanged, including incomplete closures and overlaps. No retiming, added smoothing, contact correction or biomechanics.
Timing rules · shared syllable coordination (5 November 2026)
Phases are shown on a 0–360° cycle, with at most one decimal place. Start and end delimit the gestural command window; movement and target attainment follow through the neural ramps. The same unrounded windows drive the neural decoding, phase-circle markers and time-axis guides.
The first oral onset gesture begins at 54°. All gestures in the onset cluster, including their associated glottis and velum gestures, share fixed relative timing. The onset has moved earlier while preserving every gesture duration and inter-gesture offset. Vowel and coda windows keep their established values. The table below is the current fixed rule set.
| Oral gesture | Start φ | End φ |
|---|
| Single vowel in a stressed syllable | 105.4° | 257.5° |
| Single vowel in an unstressed open syllable | 105.4° | 240.6° |
| Diphthong / vowel + vocalic R: V1 (E in lEA) | 105.4° | 223.7° |
| Diphthong / vowel + vocalic R: V2 (A in lEA) | 206.8° | 308.2° |
| Single onset consonant | 54° | 155.4° |
| Two onset consonants: first | 54° | 138.5° |
| Two onset consonants: second | 87.8° | 172.3° |
| Single coda consonant | 206.8° | 308.2° |
| Two coda consonants: first | 206.8° | 291.3° |
| Two coda consonants: second | 240.6° | 325.1° |
The unstressed open syllables with a single vowel are s@, S@, n@, ts@ and lA. lEA instead uses two separately timed vowel gestures on the same syllable oscillator: E (V1) and vocalic R (V2, A). Both retain the existing vowel amplitude and contribute to the shared vowel control parameters.
A single consonant lasts 101.4°; each member of a two-consonant onset or coda lasts 84.5°. The second consonant begins 33.8° after the first. Onset and coda are counted independently. Larger clusters are not yet covered. V1 and V2 in lEA overlap for 16.9°.
| Associated gesture | Start relative to oral start | End relative to oral end |
|---|
| Glottis opening: voiceless plosive | +33.8° | +16.9° |
| Glottis opening: voiceless fricative | +6.8° | −6.8° |
| Velum opening: nasal, default | −6.8° | +6.8° |
| Nasal directly after an obstruent in the same cluster | +6.8° | Default, unless the next rule applies |
| Nasal directly before an obstruent in the same cluster | Default, unless the previous rule applies | −6.8° |
Obstruents here include plosives and fricatives. Nasal exceptions replace the respective default offset; they are not added to it. Velum-opening gestures are created only for nasals. There is no separate obstruent velum gesture or closure rule. Glottis-opening gestures are created only for voiceless plosives and fricatives. These are the fixed relative offsets for all onset and coda positions.
Example Schrank (SraNk): Sch 54°–138.5°; A 105.4°–257.5°; glottis C1 60.8°–131.8°; R 87.8°–172.3°; ŋ 206.8°–291.3°; velum C3 200°–284.5°; K 240.6°–325.1°; glottis C4 274.4°–342°.
Current lexical scope: @ denotes schwa; ts and pf use two coordinated oral consonant gestures under the cluster rule. Short Teller is tE → lA: A represents the vocalic ER [ɐ]. Its L gesture spans 54°–155.4° and its vowel 105.4°–240.6°. Long Teller (tE → lEA) uses the same L gesture and the V1/V2 windows above. E uses the existing E target; A uses the same vocalic ER target as in lA. The mental lexicon and syllabification retain lEA; lA is an explicit realization variant on the Phono-to-Gesture side. The display is a gesture/parameter prototype, without acoustic synthesis or a dedicated affricate model.
Syllable-to-syllable timing within words
The stored lexical order defines which syllable comes next. Each syllable occurrence has its own neural oscillator, including repeated syllables. The first oscillator receives an external start signal. Every following oscillator is started by a phase-dependent signal from the immediately preceding oscillator.
| Next-syllable start signal | Rule |
|---|
| Signal begins | 162° of the preceding oscillator (phase 0.45) |
|---|
| Signal ends | 180° of the preceding oscillator (phase 0.50) |
|---|
| Command amplitude | 1.5 during this window |
|---|
The same rule is chained through the word: syllable 1 triggers syllable 2; syllable 2 triggers syllable 3; syllable 3 triggers syllable 4. Each transition uses the phase of its immediate predecessor. Ending the start signal does not stop the newly started oscillator: it continues through its own cycle.
The next oscillator is started while the preceding syllable is still running. There is no requirement to wait for its vowel or final consonants to finish, so gestures from neighbouring syllables can overlap. The new syllable’s gestures follow its own phase rules; for example, its first oral onset gesture begins at 54° of the new oscillator.
The 162°–180° interval is the prescribed trigger window. Neural decoding and synaptic filtering can delay the visible response, so the next oscillator need not visibly start exactly at 162°. This coordination rule currently applies equally to stressed and unstressed syllables and to all onset and coda structures. It was retained unchanged during the recent gesture-timing adjustments.
The original bap → gum → dip → bap sequence uses the same phase window, with the start signal additionally passing through the original SPA sequence controller. Lexical word programs pass the signal through a neural trigger ensemble to the next oscillator.
Model, playback and scope
The reference input is bap → gum → dip → bap. Word buttons select lexical syllable sequences, verified against the mental lexicon output, with the explicitly selected short-ER variant lA for Teller. Each occurrence has its own oscillator and gestures. This is an externally selected production program; it does not run the reference SPA phrase-building controller or connect automatically to the mental lexicon. The sequence retains the historical four-slot controller and oscillators. Each gesture occurrence now has its own routed ramp, including the two bap occurrences and their initial and final lip closures.
In the four-slot sequence, timing is decoded from oscillator phase, routed through inhibition/disinhibition circuits and integrated into separate gesture ramps. Both the ramp meters and section 03 group these instances by sequence slot. Their targets converge only at the shared parameter mapping in section 04. The individually selected syllables and lexical word programs connect phase decoders to gesture ramps without competing routes. Lexical word order is stored; the preceding oscillator triggers the next slot. A cascaded replay stage adds abstract agonist–antagonist ensembles. A visual-only Python port of the Trainer contours shows the resulting articulation. Articulator biomechanics and acoustic synthesis are not included.
The browser plays an actual precomputed spiking-neuron simulation. Playback speed changes viewing speed only, not the model’s speaking rate. This public version replays saved runs; no neural simulation is computed on the web server or in your browser. The Python source linked from the model overview supports local recomputation.
Source: 20_prod_4syll_map2phon_allGestures.ipynb · 2022_Syllables. Conceptual reference: Kröger (2023), JIN 22, 124; presentation slides 25–29.