A name broken apart inside one transcript event is handled — Co founder, co-founder and Cofounder. all match. A name broken across two events is not, and the agent goes silently deaf.
What it looks like
Observed on a live call, as two separate events from the same speaker moments apart:
event 1: "co-founder"
event 2: "can you tell me what is e2a"
Each is evaluated alone. The first has a wake name and no directive, so it is mention_only. The second has a directive and no wake name, so it is no_wake_name. Both are correctly rejected, and together they were a perfectly ordinary question that nobody answered.
Sketch
A bounded same-speaker buffer: when an utterance ends in a bare wake name and the same speaker continues within a short window, evaluate the pair as one. mergeBareVocatives already does exactly this reasoning within an utterance — this extends the same idea across the event boundary.
Constraints that matter more than the mechanism:
- Same speaker only. Two people, one of whom said the name, is not one sentence.
- Short window, a few seconds. Long enough for a pause, short enough that the room's next remark cannot be swept in.
- Bounded and discarded, never a growing buffer.
- The false-positive suite must not move. A missed address costs one repetition; a false one costs an interruption.
Done when
- The two-event case above triggers exactly once, not twice.
- The same input from two different speakers does not trigger.
test/asr.test.ts covers it in the spoken form, alongside the in-utterance cases.
A name broken apart inside one transcript event is handled —
Co founder,co-founderandCofounder.all match. A name broken across two events is not, and the agent goes silently deaf.What it looks like
Observed on a live call, as two separate events from the same speaker moments apart:
Each is evaluated alone. The first has a wake name and no directive, so it is
mention_only. The second has a directive and no wake name, so it isno_wake_name. Both are correctly rejected, and together they were a perfectly ordinary question that nobody answered.Sketch
A bounded same-speaker buffer: when an utterance ends in a bare wake name and the same speaker continues within a short window, evaluate the pair as one.
mergeBareVocativesalready does exactly this reasoning within an utterance — this extends the same idea across the event boundary.Constraints that matter more than the mechanism:
Done when
test/asr.test.tscovers it in the spoken form, alongside the in-utterance cases.