polyvoice 0.6.0-alpha.8

Speaker diarization library for Rust — online and offline, ONNX-powered, ecosystem-agnostic
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
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142

---
schema_version: 1
kind: module_contract
module: src/segmentation
level: subsystem
layer: algorithm
purpose: >
  Owns speaker segmentation algorithms: powerset segmentation, frame
  decoding, aggregation with Hungarian assignment, and windowed segmenter
  trait. Does NOT own embedding extraction, clustering, or VAD.
status: stable
owners:
  - polyvoice-core
workcell:
  type: leaf
  parent: ""
  children: []
  owns_paths:
    - src/segmentation/
  context_budget:
    max_files: 12
    max_source_lines: 1500
    max_contract_lines: 180
    max_readme_lines: 120
    max_todo_lines: 80
authority:
  write_policy: single_active_write_lease
  orchestrator: polyvoice-core
  read_agents: many_allowed
  migration_lease_required:
    - cross-workcell write
    - public surface migration
surface:
  - name: Segmenter
    kind: trait
    visibility: public
    contract: >
      Core segmentation trait: process audio window → speaker labels.
    proof:
      kind: unit-test
      target: src/segmentation::mod::tests
      command: cargo test --lib segmentation
  - name: PowersetSegmenter
    kind: struct
    visibility: public
    contract: >
      ONNX-backed powerset speaker segmenter (sherpa-onnx-pyannote).
    proof:
      kind: unit-test
      target: src/segmentation::mod::tests
      command: cargo test --lib segmentation
  - name: PowersetDecoder
    kind: struct
    visibility: public
    contract: >
      Decodes powerset class logits → (speaker_set, is_overlap).
    proof:
      kind: unit-test
      target: src/segmentation::decoder::tests
      command: cargo test --lib segmentation
  - name: Aggregator
    kind: struct
    visibility: public
    contract: >
      Aggregates frame-level predictions into segments using Hungarian
      assignment for speaker identity tracking across windows.
    proof:
      kind: unit-test
      target: src/segmentation::aggregator::tests
      command: cargo test --lib segmentation
  - name: FrameLabel
    kind: struct
    visibility: public
    contract: >
      Per-frame speaker set and overlap flag.
    proof:
      kind: unit-test
      target: src/segmentation::mod::tests
      command: cargo test --lib segmentation
  - name: MIN_AUDIO_SAMPLES
    kind: constant
    visibility: public
    contract: >
      Minimum audio samples required for segmentation (1600 = 0.1s at 16kHz).
    proof:
      kind: unit-test
      target: src/segmentation::mod::tests
      command: cargo test --lib segmentation
dependencies:
  internal:
    - module: types
      scope: data-shape
      reason: Confidence, TimeRange for segment timestamps.
  external: []
consumers:
  - path: .
    uses:
      - Segmenter
      - PowersetSegmenter
      - PowersetDecoder
      - Aggregator
      - FrameLabel
      - MIN_AUDIO_SAMPLES
      - polyvoice_internal
invariants:
  - id: decoder-deterministic
    rule: PowersetDecoder with identical logits always produces identical output.
    proof:
      kind: unit-test
      target: src/segmentation::decoder::tests
      command: cargo test --lib segmentation
  - id: hungarian-optimal
    rule: Aggregator Hungarian assignment minimizes total cost for speaker
      tracking across windows.
    proof:
      kind: unit-test
      target: src/segmentation::aggregator::tests
      command: cargo test --lib segmentation
verification:
  pre_change:
    - cargo test --lib segmentation
  full:
    - cargo test --lib segmentation
    - cargo test --test chaos_test --features onnx,download
    - cargo clippy --all-targets --all-features -- -D warnings
agent_policy:
  allowed_mutations:
    - Refactoring decoder internals.
    - Adding aggregation strategies.
    - Performance optimizations in frame processing.
  forbidden_mutations:
    - Changing Segmenter trait without updating all implementors.
    - Changing MIN_AUDIO_SAMPLES without checking downstream consumers.
  escalation:
    - Changes to Segmenter trait signature.
    - Changes to powerset class semantics.
---

# src/segmentation

Speaker segmentation algorithms: powerset decoder, aggregator, Hungarian
assignment, and segmenter trait.