polyvoice 0.6.9

Speaker diarization for Rust — who spoke when. ONNX-powered: Silero VAD, WeSpeaker embeddings, Pyannote segmentation, K-means/AHC clustering, overlap detection.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50

# src/der

## Purpose

Computes Diarization Error Rate (DER) between reference and hypothesis speaker
annotations. This is the single source of DER truth for the polyvoice project.

Uses frame-based evaluation at 10ms resolution with a configurable forgiveness
collar around reference boundaries. Speaker IDs are mapped optimally via greedy
1-to-1 matching on co-occurrence counts.

## Surfaces

- `compute_der(reference, hypothesis, collar) -> DerResult`
- `DerResult` — struct with `der`, `miss_rate`, `false_alarm_rate`, `confusion_rate`, `total_speech`
- `compute_der_from_rttm(reference, hypothesis, collar) -> DerResult` — convenience wrapper for string-labeled references

## Dependencies

- `types` — `SpeakerTurn`, `TimeRange`, `SpeakerId`

## Invariants

- DER is always in [0.0, 1.0]
- Identical ref/hyp → DER == 0
- Empty reference → DER == 0
- miss + false_alarm + confusion == der (within f64 rounding)
- Collar reduces or preserves DER for boundary errors

## Verification

```bash
# Fast pre-change check
cargo test --lib der
cargo test --test property_der_test

# Full verification
cargo test --lib der
cargo test --test property_der_test
cargo test --test der_baseline_test
cargo clippy --all-targets --all-features -- -D warnings
```

## Notes

- The `benches/der_ami.rs` benchmark currently duplicates a simplified DER
  implementation instead of using this module. See TODO.md.
- Standard collar value is 0.25s (per NIST evaluation protocol).
- Greedy mapping is not globally optimal but is fast and sufficient for
  diarization evaluation.