Add dependency-free diarization error rate (DER) scoring

[alexkroman]

Adds a new aai_cli.core.der module that implements NIST diarization error rate (DER) scoring in pure Python, with no external dependencies beyond the standard library.

Summary

DER is the standard metric for evaluating speaker diarization — it measures the fraction of reference speech time that is misattributed to the wrong speaker. This implementation complements the existing WER (word error rate) scoring in aai_cli.core.wer by providing speaker attribution accuracy.

Key Changes

• New module aai_cli/core/der.py: Implements DER scoring with:

  • Segment dataclass: represents a span of speech attributed to a speaker
  • Score dataclass: holds the three NIST error components (missed, false_alarm, confusion) plus reference speech time, with computed properties for total errors and DER rate
  • score() function: computes DER by partitioning the timeline at segment boundaries, tallying errors per atomic interval, and finding the optimal one-to-one speaker mapping via exhaustive search (feasible because diarization files have few speakers)
  • pooled() function: aggregates scores across multiple files for corpus-level DER
  • Helper functions for timeline analysis, speaker mapping, and cooccurrence tracking

• Comprehensive test suite tests/test_der.py: 14 tests covering:

  • Immutability of value types
  • Perfect diarization (zero error)
  • Speaker label remapping (labels are arbitrary, optimal mapping recovers correct attribution)
  • Missed speech, false alarms, and speaker confusion scenarios
  • Overlapping speakers (counted per speaker, not wall-clock time)
  • Optimal vs. greedy speaker assignment
  • Corpus-level pooling

Implementation Details

• No external dependencies: Uses only Python stdlib (itertools, dataclasses, collections.abc)
• Optimal speaker mapping: Exhaustive permutation search over speaker pairs (factorial complexity is acceptable for typical diarization speaker counts of 2–10)
• Timeline partitioning: Segments are split at every boundary so each atomic interval has a fixed set of active speakers, enabling accurate per-interval error tallying
• Weighted by speaker-time: Reference speech time is counted per concurrent speaker, so overlapping speech is properly attributed
• Immutable value types: Both Segment and Score are frozen dataclasses for safety

https://claude.ai/code/session_011qBbDEWrtpPVQVaVKAHnMx