What does this PR do ?

_DetectionErrorRateAccumulator (used by vad_tune_threshold_on_dev and frame_vad_eval_detection_error) replaced pyannote's DetectionErrorRate in #15632, but it undercounts false alarms. With no UEM, md-eval scores only the reference's extent, so speech predicted before the first or after the last reference segment is never counted. pyannote scored the union of the reference and hypothesis extents.

Example: reference [2, 4], hypothesis [0, 1] + [2, 4] + [5, 6]. pyannote gives FA = 2.0 s (DetER 100%). main gives FA = 0 (DetER 0%).

I compared both on 1,600 random reference/hypothesis pairs, with and without overlapping segments. main disagreed with pyannote on false alarm in 1,066 of them. With this change, every pair whose reference has speech matches.

Collection: ASR

Changelog

• When no UEM is given, score from the earliest start to the latest end across reference and hypothesis.
• Add test_detection_error_rate_counts_false_alarm_outside_reference. It fails on main and passes here.

One difference remains, and it's out of scope here: a file with no reference speech is skipped by evaluate(), so its false alarms still don't count.

Checks run

• pytest --cpu -m "not pleasefixme" tests/collections/asr/utils/test_vad_utils_asr.py tests/collections/speaker_tasks/utils/: 700 passed, 37 skipped
• pre-commit run on the changed files

No documentation change needed.

PR Type:

• Bugfix