Refine back-channel classification when strict text lookup is empty

[julien731]

Background

While testing PR #59 / issue #51 against the Nose Tea Pre-kickoff Meeting (~80 min, 6 speakers), interaction analysis produced 190 interruption events. Spot-checking revealed that 28 of the 190 (~15%) are short (≤ 1.5s) and have ack-y context strings ("Yeah.", "Okay.", "Yes.", "Yeah, yeah, yeah."), suggesting they should have been classified as back-channel overlap rather than interruption.

Examples (real meeting 955eeb48-…):

t (s)	overlap dur	a → b	context
136.16	0.71	Pat → Carlos	"Yeah."
192.46	0.47	Pat → Julien	"Okay."
332.89	1.15	Julien → Got	"Yes."
1481.01	1.01	Carlos → Julien	"Yeah."
1708.22	1.13	Pat → Carlos	"Yeah, yeah, yeah."

Root cause

_detect_overlaps in backend/services/interaction_analyzer.py classifies an overlap as a back-channel only when _is_backchannel(duration_b, text_b_strict) returns True. The strict text lookup (introduced in af44ac7 to avoid mis-attributing nearby segments) returns \"\" when there is no transcript segment from speaker B that strictly contains the midpoint of B's PyAnnote turn.

When the strict lookup is empty, _is_backchannel only catches the case where B's turn is ≤ 0.5s AND has no text. For 0.5s < duration_b ≤ 1.5s, the lexical check on the empty string returns False and the event falls through to INTERRUPTION.

The displayed context field of those events comes from a lenient lookup (nearest-edge segment), which can pull a Carlos "Yeah." from minutes away — making the JSON output misleading: the displayed text is not actually B's utterance at that timestamp.

Why this happens in practice

PyAnnote sometimes emits a sub-1.5s diarization turn for a speaker for whom WhisperX did not produce a corresponding speaker-tagged transcript segment (typical for very short vocalizations during another speaker's run). The current rule is conservative: "no strict text → assume substantive interruption." In this dataset that produces ~15% false-positive interruptions for what subjectively look like back-channels.

Proposal (to discuss)

Extend the back-channel heuristic so that, when the strict text lookup is empty:

• Treat turns with duration_b ≤ BACKCHANNEL_MAX_DURATION (1.5s) as back-channel iff the lenient lookup at the midpoint resolves a segment that is (a) within ±X seconds of start_b/end_b AND (b) lexically a back-channel. Otherwise default to interruption.

Tune X empirically against the existing fixture corpus + this real meeting. Add a test fixture covering the "PyAnnote turn with no strict speaker-B segment" path.

Independently, consider revising the event's context field so we don't emit lenient text that misleads the reader — either label it (context_source: 'lenient' | 'strict') or omit it when no strict text was found.

Out of scope

• Any UI rendering changes (Phase 4).
• The dominance / pause / per-segment annotation paths — those tested clean against AC.

Verification

Re-run audio analysis on 955eeb48-b55e-4e8f-9d1a-89fcb99ca609 (or the fixture corpus) and confirm:

• Total interruption count drops by ≈ the number of obvious back-channel false-positives
• Newly-reclassified events appear under overlap
• AC1/AC3 tests still pass

Priority

Optional / nice-to-have. The current behaviour is conservative-by-design; this is a refinement, not a defect.

Related

• PR [#51] Add interaction pattern detection (interruptions, hesitations, dominance) #59 — [As a user, I can have interaction patterns detected so that interruptions and turn-taking dynamics are captured #51] Add interaction pattern detection
• Issue As a user, I can have interaction patterns detected so that interruptions and turn-taking dynamics are captured #51 — interaction pattern detection (parent)
• Commit af44ac7 — strict-overlap text lookup hardening