No neural networks. No recordings. No training data. No runtime dependencies. A digital waveguide that produces human-quality speech from mathematics alone. Formants emerge from geometry, not lookup tables.
Sound is produced by simulating the physical vocal tract. No statistical models. No neural inference. Pure waveguide physics.
Liljencrants-Fant two-mass fold model generates the raw excitation signal. No samples, no recordings — pure waveform physics from first principles.
10-section digital waveguide models the human vocal tract. Cross-sectional areas shape formant frequencies. Change the geometry, change the voice.
Lip radiation model applies the final acoustic transform. Natural-sounding speech computed entirely from first principles. Real-time on embedded hardware.
The same 2 MB kernel powers every application. No cloud. No network. Runs on embedded hardware in real time.
Natural speech restoration for laryngectomy patients. Real-time synthesis from a 2 MB kernel running on embedded hardware. No cloud connection required.
Cochlear waveguide processing with Hopf amplifier and binaural spatial audio. Physics-based hearing aid kernel in 2 MB with zero latency.
Extensible waveguide framework for tactile and spatial sensing prosthetic devices. The same physics kernel adapted for non-auditory sensory input.
Deterministic voice synthesis and recognition for secure communications. No external dependencies, air-gapped capable, tamper-proof.
The theoretical foundations behind the waveguide engine. All peer-reviewable with permanent DOIs.
Voice production and auditory perception are mathematical duals. The same kernel powers both — one physics framework, two directions. Production (W) and perception (W†) form an adjoint pair on a 2-simplex. One architecture for all sound.