Aloud
Lets someone book a Greek public-health appointment by speaking: streaming speech-to-text feeds a deterministic, fully tested engine that fills the real search form. It runs on top of the FindDoctors aggregator; two more demos cover legal intake and restaurant ordering. The orchestration behind the LLM front end is scripted code that can be tested, rather than another prompt.
The problem Link to heading
Voice interfaces usually stop at dictation. Turning “βρες μου καρδιολόγο κοντά μου την άλλη εβδομάδα” into an actual search, with slots, areas, and dates resolved correctly, needs a deterministic engine behind the microphone.
What I built Link to heading
Built a scripted, testable orchestration engine in TypeScript first, then ported it to Go for production. The Go port is verified against the TypeScript reference with a 32-scenario golden harness that produces byte-identical output.
Implementation Link to heading
- Deterministic Voice Engine: Intent extraction, specialty/area/date resolution, and form merging run scripted and fully testable, with no API key required.
- Go Port with Golden Parity: voice-core reimplements the TS engine in Go; parity is enforced by ported unit tests plus a 32-scenario golden harness, byte-identical to the reference.
- Resilient Sessions: Per-session actors with ordered turns, dedupe, bounded event logs, and reconnect-resume over a versioned WebSocket contract; go test -race clean.
- Real Data Behind the Voice: Plugs into the FindDoctors Go aggregator and soniox-core streaming STT, so the demo searches real availability.
- Optional LLM Layer: Claude can rephrase assistant replies into more natural Greek without touching the deterministic flow.