← Model overviewNEF · SPA · NENGO / MODEL 0.15Research ↗

INTENDED CONCEPTS → ONGOING ARTICULATION

From Concept
to Articulation.

Follow intended concepts through lexical retrieval and syllable memory.
Watch preparation and articulation advance on one model timeline.

CONCEPT → LEMMA → PHONOLOGY

Ordered conceptual intentions
Neural mental lexicon
Neural lexical control · model 0.15

Eight supported concepts · 613 lexical entries in the neural pathway.

01 Prepared conceptual plan

Checked recordings · model 0.15

Loading the conceptual plan…

Each recording contains 48 word occurrences from a limited repertoire, not 48 different words. The externally supplied concept list is executed through the neural lexicon. The panels below show its ongoing retrieval and articulation.

Model 0.15 · recorded examples

Neural target mixture

Loading the recording catalogue…

The eight checked 48-word recordings are ready to play without recomputation. They show the neural-mixture output at its original model time. Use the timeline, 5-ms steps or slower playback to inspect articulation. This public viewer plays saved causal online simulations; it does not run a new Nengo simulation in your browser. A recording is downloaded when opened.

Measured timing for these examples

Loading the measured comparisons…

Eight complete corpus runs, one whole-model seed. All 384 words, 608 syllables and 2,496 gestures passed the recorded audits. Rates use the first syllable onset to the final syllable offset. Download validation data

0.000 sNo recording loaded

One shared model timeline · playback of the computed online run Download viewer data (.json.gz)

03 Phonological sequence memory

—

ReadySelectedAccepted / clearingLoadingR = read address · W = write address

Each numbered place is a separate recurrent SPA state. Positions are reused cyclically; repeated syllables retain distinct occurrences. The fill bar shows the largest decoded pointer coordinate.

PREPARED NEXT · P_next

—Waiting for input
→

MOTOR HANDOVER

Waiting for a prepared syllable

Neural read and write addresses are initializing.

04 Ongoing execution

Overlapping syllable occurrences

No active oscillators yet.

Gesture activation

Current recurrent ramp states

05 Articulation

Same model time

Glottal aperture
superior view

Current gesture ramps pass through the spiking neural target mixture before driving positive and negative opponent populations. The displayed contours come from that neural output, with no additional smoothing or time shift. Geometry and gestures share the same forward run.

06 Sequence events

Events at or before the selected time

What model 0.15 implements

An external list specifies the intended concepts. Neural address, request and handover circuits execute the list through the 512-dimensional concept → lemma → phonology pathway with 613 lexical entries. A neural association supplies the recurrent eight-dimensional P_word identity. Stored word/position connections and neural offer/acknowledgement circuits deliver syllables incrementally to the ten-place phonological memory.

Memory addressing, preparation, motor start, oscillator acknowledgement and context release are neural. The first syllable can be prepared during lexical retrieval; following preparation overlaps ongoing execution. Initial lead (two or four syllables), refill target (four ready syllables) and capacity (ten places) are separate settings. Every nonfinal motor start requires the following syllable ready. A whole word need not fit into free places at once.

The target mixture uses 62,800 LIF neurons for bounded ramp activation, fixed target projections, normalization and neutral remainder. Both opponent directions advance in the same persistent 1-ms simulation; frames are recorded every 5 ms. Playback is a view of that causal online computation, not its execution speed.

Lexical acceptance and P_word write/clear control are neural. The confirmed successor can cue the next concept before address copy-back finishes; the held word stays protected. Motor reuse requires confirmed capture of the guard before retirement, with the nominal 200-ms guard unchanged. Parameter gains and vocal-tract geometry remain the numerical output boundary. The intended list is not a neural sentence generator. There is no acoustic synthesis. At the current tempo, rates are 1.856–2.337 syllables/s. At the offline gesture threshold 0.10, only the mixed recordings contain short gaps (10 ms with lead two, 5 ms with lead four); these are not measured acoustic pauses. Higher speaking rates have not been tested. Visual inspection complements the recorded functional tests.