VoxTeX

Speech-to-math input for browser assessments — a client-side-only proof of concept.

Targets accessibility for students with mobility impairments: dictate a math expression, hear it read back, confirm hands-free, and get symbolic correct/incorrect feedback. No backend, no inference APIs — ASR, parsing, rendering, read-back, and answer checking all run locally in the browser.

Pipeline

mic (MediaRecorder → 16 kHz mono PCM)
→ Whisper tiny.en (Transformers.js, Web Worker, WebGPU with WASM fallback)
→ spoken-math grammar parser (deterministic recursive descent → AST → LaTeX)
→ KaTeX render
→ speech-rule-engine (ClearSpeak) → speechSynthesis read-back
→ "yes / no" voice confirmation through the same ASR stream
→ @cortex-js/compute-engine symbolic equivalence check

2 times the quantity x plus 3 end quantity is judged correct against an expected answer of 2x+6 — answers are compared symbolically, not textually.

Repo layout

Path	What
packages/spoken-math-parser	The centerpiece: phrase → LaTeX parser. Pure TypeScript, no DOM — reusable by future input lanes (ink, photo OCR). 165 Vitest cases.
packages/voice-math-input	The embeddable SDK: mic → Whisper worker → grammar → LaTeX as one headless package, plus a <voice-math-input> Web Component. See INTEGRATION.md for consuming it from webpack/Vite apps or via script tag, and for self-hosting all assets on a private CDN.
apps/web	Vite + React demo: one assessment-item page with mic button, live transcript, rendered math, spoken read-back, feedback, and a debug panel. Consumes the built SDK package, doubling as the reference integration.

Run it

npm install
npm test# parser unit tests (75 cases)
npm run dev # http://localhost:5173
npm run build # production build

Requirements: a Chromium-based browser is the happy path (WebGPU). Any browser with WASM works via the fallback. First click on Enable voice input downloads the Whisper tiny.en weights (~40 MB) from the Hugging Face Hub — that's a one-time static file download, cached by the browser; inference never leaves the machine. Everything else (SRE mathmaps, KaTeX fonts) is served from this origin.

The ambiguity rule (the important design decision)

"one over x plus two" is ambiguous. VoxTeX resolves it deterministically:

• over binds tightly → one over x plus two = ⅟ₓ + 2
• Widen with the quantity … end quantity → one over the quantity x plus two = 1/(x+2)
• Or use all over, which splits everything said so far → x plus one all over two = (x+1)/2

The read-back loop makes the rule self-correcting: the student hears which interpretation was chosen ("the fraction 1 over x, end fraction, plus 2") and can say "no" to retry. Full precedence table and corollaries: packages/spoken-math-parser/README.md.

Hands-free flow

1. Press the mic button (keyboard accessible; all status changes announced via aria-live). Speak: "two x plus six". Recording auto-stops ~1.5 s after you stop talking (energy-based end-pointing with a calibrated noise floor; the button still works as a manual stop, hard cap 30 s).
2. The transcript is parsed and rendered with KaTeX; SRE reads it back and asks "Say yes to submit, or no to try again."
3. A confirmation window listens through the same Whisper worker, end-pointed the same way (stops ~0.9 s after you answer). yes/confirm/submit → symbolic check + spoken verdict; no/try again → dictation restarts automatically. Silence or an unclear reply triggers one spoken re-prompt, then falls back to on-screen Confirm / Try again buttons.

Typed input ("type the spoken phrase") exercises the identical pipeline without the mic — useful for testing and as another access lane.

Design notes & limitations (v1)

• Why not the Web Speech API: in Chrome it ships audio to Google servers — not client-side, and unusable for secure assessments. Whisper-in-worker keeps audio on device.
• Whisper sometimes emits symbols (2x + 3, 1/2); the tokenizer accepts both words and symbols.
• Vocabulary now covers fraction words, trig, logs, integrals, summations, derivatives (including prime notation), ordinal powers ("x to the fourth"), and common Greek letters (see parser README). Notable gaps: limits, products, multi-letter identifiers.
• Compute Engine grades finite sums and symbolic derivatives (the sample items use both), but not definite integrals.
• Definite integrals parse and render, but the Compute Engine reports "unknown" when asked to equate them to a closed form, so integrals aren't used in the graded sample items.
• End-pointing is a simple RMS-energy VAD (calibrated noise floor) — good enough for quiet rooms; a spectral VAD would be the next step.
• Personalization (opt-in): the SDK learns per-user mishearing corrections from confirm/reject feedback (constrained so valid math can never be rewritten) and persists a mic noise-floor profile — all in local browser storage, with inspect/export/clear APIs. See INTEGRATION.md §6.
• speechSynthesis voices vary by OS; the math speech text itself is deterministic (SRE ClearSpeak, with the parser's own read-back as fallback).