Voice control layer for transit operations · AssemblyAI → GTFS-RT

Speech is a draft.
Truth is versioned.

Who: Transit controllers dictating live detours and skipped stops.
Without it: “Wait, keep Cumberland” becomes a second, conflicting alert.
ERRATA: Repairs the same change_id as the next revision.

Voice proposes. Humans commit. Speech only ever produces an outlined draft. Apply puts it on air; a hash-checked Commit makes it final.

LIVE CAPTION AssemblyAI Universal-3.5 Pro Realtime

DISCONNECTED

Mic
Caption
Draft no effect
Apply human
On air

Hearing now draft · not applied

No speech buffered.

ERRATA replies

Ready when you are

Start the microphone. ERRATA will greet you, interpret each completed turn without mutating canonical state, and tell you what it needs next.

Next: connect the microphone.

No microphone?

Type it — same parser, same reducer

Voice engine & guidance

Voice engine CHECKING…

AI33 voice Browser fallback

Connect — · Interpret — · Voice —

Browser microphone → AssemblyAI Universal-3.5 Pro Realtime → non-mutating ERRATA interpretation → explicit human ForceEndpoint/apply → the same parser, validators and reducer used by direct entry.

02 Timeline

One change_id. Every frame is a revision.

Corrections cut a new frame on the same timeline. Dropped frames never air — zero canonical effect, hash unchanged.

03 Commit

Nothing is final until a human signs the current hash.

04 Feed

What goes out is decoded, not described.

05 Ledger

Everything an auditor would ask for.

Demo control

Reset staged change

Returns to the seeded service date/start time and revision 1.