Concept2Articulation Pathway
How can the next word be retrieved while earlier syllables are already being spoken? This coupled NEF-SPA model connects an intended sequence of concepts to a neural mental lexicon, a phonological sequence memory and ongoing articulatory execution. Its central feature is temporal overlap: lexical retrieval, syllable preparation and motor execution advance together on one model timeline.
From Concept to Articulation — Interactive Lab →
Explore eight checked recordings, compare two or four syllables of initial preparation, and follow the concepts, buffer contents, oscillators, gestures and articulatory movements. No local installation is needed.
The public lab plays saved simulations. New neural calculations are performed in the local Python/Nengo version. Opening a recording downloads its compressed trajectory.
1. From conceptual intentions to movement
An externally supplied enumeration plan specifies the intended concepts. Neural control advances through this plan and retrieves each candidate through concept → lemma → phonological form. The retrieved word identity is held in a recurrent register, P_word, while its stored syllable sequence is delivered incrementally to phonological memory.
Input boundary: externally specified concept sequence
Concept → lemma → phonology
Protected word register P_word
Cyclic write and read addresses
Prepared next syllable P_next
Neural target mixture
Positive / negative opponents
Coupled control: supply demand and free memory places regulate lexical delivery; readiness, oscillator phase and available motor contexts regulate the next syllable start.
Output boundary: numerical articulatory parameters → vocal-tract geometry
Each phonological memory place retains a syllable occurrence. Separate places preserve order and distinguish repetitions of the same syllable. Neural readout prepares the next occurrence for a syllable oscillator. Its associated gesture program then drives recurrent ramps, a neural target mixture and opponent populations, whose outputs determine articulatory parameters.
2. Preparation overlaps execution
The ten-place memory does not have to be full before speaking starts. The examples start after either two or four syllables are ready. Each subsequent nonfinal start requires the following syllable already to be available. A refill target of four ready syllables regulates further supply; capacity, initial lead and refill demand are different settings.
Read and write addresses cycle through the memory. Places are released and reused after the neural handover and clearing steps. Retrieval of a following concept can overlap delivery of the current word, and preparation of a following syllable can overlap ongoing articulation. The next concept therefore does not wait for the preceding word’s articulation to end.
Neural processing time does not necessarily create a pause. A process can take time while earlier material continues to be executed. Delays become disruptive when the required result is still unavailable at its motor deadline. The viewer distinguishes waiting for a start from an interval without active gestures.
3. What to look for in the lab
- Start with Mixed · 48 words · lead 2. Press Play, or use the timeline, 5-ms steps and “Go to model time”. Playback defaults to one model second per second; slower playback supports inspection.
- Compare the concept/lemma/phonology panels with the protected P_word register. At about 0.800 s in this recording, the lexicon is already preparing Kanne while the word register still holds Tasse.
- Follow the read/write markers in the ten memory places, P_next, and the active syllable oscillators. These panels show preparation advancing alongside execution.
- Compare the corresponding lead 4 recording. The timing table and rate summary refer to the actual recorded runs. Gesture bars and the articulatory view share their original time axis.
4. Checked recordings and current scope
The mixed, monosyllabic, disyllabic and repetition profiles passed the recorded stability, timing and articulation audits. The lexical network contains 613 entries, while this coupled version has prepared motor programs for eight distinct words: Tasse, Tasche, Kanne, Katze, Teller, Topf, Stuhl and Schrank. Each 48-word sequence repeats a subset of this repertoire. Teller with long R is a pronunciation variant of the same lexical word. These are not tests of 48 different words.
The underlying online model uses a persistent neural simulation with 1-ms integration and simultaneous positive and negative opponent channels; displayed frames were recorded every 5 ms. No whole-utterance opponent replay is used in generating these trajectories. Browser playback speed is independent of the computation time needed to simulate the network.
Rates use the interval from the first syllable’s first gesture onset to the last syllable’s last gesture offset, with an offline activity criterion of 0.10. Initial preparation and the technical recording tail are excluded; internal gaps remain included. At this criterion, the two mixed recordings contain short all-gesture gaps of 10 ms (lead 2) and 5 ms (lead 4). These are model measurements, not measured acoustic pauses.
Limits of this demonstration. The results use one whole-model seed and a relatively slow fixed tempo. Higher speaking rates and a larger set of distinct words remain to be tested. The model reads stored word-to-syllable sequences; it does not generate grammatical sentences or perform general resyllabification. Opponent populations provide an abstract muscle model. Numerical geometry is the output boundary, and there is no acoustic synthesis or validation against physiological recordings.