ʋ voksa

Problem

Lojban is a spoken language with no voice. Text-to-speech for a constructed, phonetically-regular language shouldn't need a neural net or a server — but it also shouldn't sound like a 1980s formant synth read from a table nobody agreed on.

Approach

voksa is a rule-based Klatt-style formant synthesizer written in pure Rust, compiled to a ~42 KB, zero-import WebAssembly module that runs the whole synthesis on the browser's main thread (< 100 ms per sentence) and plays through a tiny AudioWorklet — no network after load. The interesting part is tuning: the CLL defines what an attitudinal means but never how it sounds, so the prosody, naturalness, and per-phoneme voice tables are voksa's invention. The tuning console (a Dioxus component) exposes all 449 parameters with Lojban-labeled controls and English glosses, a live phonetic-analysis line, and an export/load loop: fiddle the voice, listen, and share a config JSON back to the community that replays bit-identically on the native CLI.

Outcome

Offline Lojban speech in any browser, and a way for the Lojban community to converge on how the language should actually sound — by ear, together, with every tuning reproducible from a small JSON file.

Try it live at /voksa — the real engine compiled to WebAssembly, the full tuning console in your browser. Nothing leaves your tab.