From syllables to articulation
Fifty German words. Ten five-word chunks. Follow the neural timing, gesture activations and articulatory movements on one shared timeline.
01 Choose a recording
50 words · 92 syllable occurrencesSelecting a word opens its five-word recording and moves to that word’s first observed syllable activation. Neighbouring words remain in context; no reset or inserted pause occurs inside a chunk.
Loading P2G-02 recordings…
02 Syllable oscillators
Select a syllable to inspect its gesturesEach occurrence has its own oscillator. The dot shows its decoded state; a quiet oscillator sits near the centre. Main stress: 1.2 Hz; unstressed: 1.5 Hz. Playback speed does not change these model frequencies.
03 Gesture activation
0.000–2.000 s · shared two-second viewAll five panels stay open and share a moving two-second view. After the first second, the time marker stays centred while the view follows continuously; the first and last windows stay fixed at the recording boundaries. Use the shared timeline above or click a plot. Gesture curves appear for the currently active words.
Selected syllable: current values and prescribed phases
Curve visibility spans the first detected start signal or gesture to the last detected gesture of each word, including gaps between its syllables. During word overlaps, both panels show their curves. All five panels and their time axes remain visible. Dashed grey curves show syllable start signals. This display filter uses the recording’s activity intervals; it does not alter the recorded curves.
| Gesture | Start | Target | End |
|---|
04 Articulatory movement
Visual contour model · facing left. Current online opponent outputs drive tongue, lips and velum. Sch uses one apical gesture with postalveolar tongue-tip position; lip rounding is supplied by vowels. No acoustic synthesis.
05 Integrated articulatory parameters
Positive and negative neural populationsBars show decoded positive and negative opponent activity. The signed value is their unscaled difference; “model” includes the established geometry gains and offsets.
06 Recording notes
Observed activation issues
Syllable transitions: measured phase alignment
| Transition | Previous end (s) | Next 108.22° (s) | Difference (ms) | Measurement |
|---|
Differences use the first forward crossing. Where the decoded phase crosses a reference more than once, the measurement column gives the full candidate range. Positive means the successor reaches its reference phase later; negative means earlier. These are decoded oscillator phases, not measured biomechanical target times. A physiological acceptance tolerance has not been established.
Model, timing and scope
P2G-02 extends P2G-01 with the approved fifty-word corpus, short/long vowels, two-vowel nuclei, stress-dependent oscillator frequencies and frequency-dependent successor starts. The next oscillator is driven by a neural phase-trigger ensemble; the externally supplied phonological sequence does not use the Concept-to-Articulation lexical pathway.
Prescribed successor start phase = previous primary gesture end − (previous frequency / successor frequency) × 108.22°. The neural start window is 18° wide. Neural approximation and filtering can produce deviations; the diagrams show the recorded output, not an idealized trajectory.
In this revision, oscillator feedback compensates for the finite integration step during phase progression. The larger startup impulse is retained. The original gesture and start filters are retained after separate diagnostic checks.
One persistent simulator per chunk integrates oscillators, gesture populations and both opponent directions at 1 ms. Output is recorded every 5 ms. Target mixing and contour geometry use the established numerical adapter. Contours are exported from simultaneous neural outputs, without a second neural simulation, retiming or extra temporal smoothing. This is a visual model, not a biomechanical or acoustic validation.
The public website plays saved neural recordings. The local Python version additionally offers “Recompute this chunk”. These are implementation diagnostics using seed 11 and opponent seed 23, not a validated corpus study.