polyvoice 0.14.0

Speaker diarization for Rust — who spoke when. ONNX path optional: default features are empty (ort-free BYO-embedder core); enable onnx for Silero VAD, WeSpeaker embeddings, and Pyannote segmentation.
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
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310

schema = "polyvoice-models-v2"

# Stage-scoped version aliases (Deepgram diarize_model=latest|v1 pattern).
# Resolution is logged so DER reports record the pin that was actually used.
[aliases.segmenter]
latest = "powerset_fp32"
v1 = "powerset_fp32"

[aliases.embedder]
latest = "wespeaker_resnet34"
v1 = "wespeaker_resnet34"

[aliases.vad]
latest = "silero_vad"
v1 = "silero_vad"

# Shipping profiles. Both default to the proven FP32 pair: the recalibrated
# INT8 pair is much faster (~7x embedding stage, ~2x segmentation stage) but
# regresses DER on AMI meetings (+2.0/+2.8 pp), so it ships as the opt-in
# `fast` profile. The FP32 models stay available via their model ids for
# direct ModelRegistry::ensure() callers.
[profiles.mobile]
segmenter = "powerset_fp32"
embedder  = "wespeaker_resnet34"

[profiles.balanced]
segmenter = "powerset_fp32"
embedder  = "wespeaker_resnet34"

# Opt-in speed profile: recalibrated INT8 pair (powerset weights-only dynamic
# + ResNet34 static QDQ). At DER parity on VoxConverse-style audio (+0.2 pp
# on the dev split); expect ~+2 pp DER on AMI-style meeting audio.
[profiles.fast]
segmenter = "powerset_int8"
embedder  = "resnet34_int8"

# Legacy v0.5 entries — kept for back-compat callers that pass the model id
# directly to ModelRegistry::ensure(). Profiles do not reference them anymore.
[models.silero_vad]
url      = "https://github.com/snakers4/silero-vad/raw/master/src/silero_vad/data/silero_vad.onnx"
sha256   = "1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3"
size     = 2327524
filename = "silero_vad.onnx"
license  = "MIT"
license_url = "https://github.com/snakers4/silero-vad/blob/master/LICENSE"
provenance = "snakers4/silero-vad upstream ONNX"
adapter_type = "silero"
version  = "1.0"
sample_rate = 16000
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhZU2jujLiO9QWKC7WPSGEE1crTXcP/4t+FGlbFrJ9+JLTVuj7Om/zXdAo1Aak/nvQo3/7xXev41Qn10+VSda/wE=
trusted comment: polyvoice v0.6.0-alpha.3 | silero_vad.onnx
ECe8pg8lcsO5MxlmAjaLRUIFC2t5TRt8gGrl3boQO7PiVJFZkzgFlgI74YH9T1Dp0bKgaMBJ0kSBIJpmyZVmDw==
'''

[models.wespeaker_resnet34]
url      = "https://huggingface.co/Wespeaker/wespeaker-voxceleb-resnet34/resolve/main/voxceleb_resnet34.onnx?download=true"
sha256   = "9fea6516d7ad6bf0a76c7689f5a49b65d330fad6dde96c91bb4435ffbfe056a1"
size     = 26534127
filename = "wespeaker_resnet34.onnx"
license  = "Apache-2.0"
license_url = "https://huggingface.co/Wespeaker/wespeaker-voxceleb-resnet34"
provenance = "Wespeaker/wespeaker-voxceleb-resnet34 Hugging Face"
adapter_type = "wespeaker-resnet34"
version  = "1.0"
sample_rate = 16000
embedding_dim = 256
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhf99QhdU+yew+XOYA3tx+dixo9cqxxz0Y7xlIOiiLhFTFKbsKeiH5OkiGn9GPbzP8TwGGLjKGLVRsOdFHSueOAs=
trusted comment: polyvoice v0.6.0-alpha.3 | wespeaker_resnet34.onnx
BzPwRAu4i4ABLmMMqmgv++OWy+3tbmdf9FrCIgXtB/zfdXXwckWIQE7vcCIfpuLPS0BUWDbkbbV6n3mugnlCBw==
'''

[models.powerset_fp32]
url      = "https://huggingface.co/csukuangfj/sherpa-onnx-pyannote-segmentation-3-0/resolve/main/model.onnx"
sha256   = "220ad67ca923bef2fa91f2390c786097bf305bceb5e261d4af67b38e938e1079"
size     = 5992913
filename = "powerset_fp32.onnx"
license  = "MIT"
license_url = "https://github.com/k2-fsa/sherpa-onnx"
provenance = "sherpa-onnx-pyannote-segmentation-3-0 (pyannote/segmentation-3.0)"
adapter_type = "powerset-v1"
version  = "3.0"
sample_rate = 16000
window_secs = 10.0
hop_secs = 2.0
num_speakers = 3
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhQ10AL95cKcDAXudXyE1DdH7VfQpci6E/PZHNlI6W19DEjsqPi8tZ7GC8PZkaHeRJ4ZnjAKTQCvkRWYoByTjuAk=
trusted comment: polyvoice v0.6.0-alpha.3 | powerset_fp32.onnx
e33pe01miZQKvp1AoCQcv6Oa3vVmxOBcxNiOkasmsCxhRq5ix1uqMWjah8IB6YieUjvHYj4hd9j1OH6wPAm/AA==
'''

[models.cam_pp_fp32]
url      = "https://huggingface.co/Wespeaker/wespeaker-voxceleb-campplus/resolve/main/voxceleb_CAM%2B%2B.onnx?download=true"
sha256   = "b50810498b5bcf5773d086f6993d344476bd0c88b566a41e8d801aaf8461efad"
size     = 29292449
filename = "cam_pp_fp32.onnx"
license  = "Apache-2.0"
license_url = "https://huggingface.co/Wespeaker/wespeaker-voxceleb-campplus"
provenance = "Wespeaker/wespeaker-voxceleb-campplus Hugging Face"
adapter_type = "cam++"
version  = "1.0"
sample_rate = 16000
embedding_dim = 512
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhZxjhT1+WXcKvuaszc4e4LHkeeySB9tzOhQLjB23OeDyBKfzQWQdfp1JfLylKJ3fvlH92eix78kBHpFruwFCFAc=
trusted comment: polyvoice v0.6.0-alpha.3 | cam_pp_fp32.onnx
pnLedo8TIV+K14fAyL54r1d9mWm1rN18C50BA/QQclyfP01tOaJVbxcJB+GZ4+AA/fluvTDJewlPn1gMo33wBA==
'''

# v1.0 INT8 artifacts, recalibrated against VoxConverse-dev:
# - powerset_int8: weights-only dynamic quantization (QInt8, per-channel) —
#   static QDQ is not viable for this recurrent model (int8 activations on the
#   LSTM path destroy posteriors even when every compute op is excluded).
#   No calibration data.
# - resnet34_int8: static QDQ (QInt8, per-channel), calibrated on 1819 real
#   fbank+CMVN speech windows (0.5-5 s RTTM segments, VoxConverse-dev), with
#   the statistics-pooling tail and the input conv kept in fp32.
# - cam_pp_int8: unchanged from the previous calibration; not referenced by
#   any shipping profile.
[models.powerset_int8]
url      = "https://github.com/ekhodzitsky/polyvoice/releases/download/models-int8-v2/powerset_int8.onnx"
sha256   = "175896d26f639933cd86906d2dd3e6796eddb23c1f719925a3949052da76183b"
size     = 1649256
filename = "powerset_int8.onnx"
calibration = "weights-only dynamic quant, no calibration data"
license  = "MIT"
license_url = "https://github.com/k2-fsa/sherpa-onnx"
provenance = "INT8 quant of sherpa-onnx-pyannote-segmentation-3-0"
adapter_type = "powerset-v1"
version  = "3.0-int8"
sample_rate = 16000
window_secs = 10.0
hop_secs = 2.0
num_speakers = 3
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhcR4tl4/adcGnamCx9bdqn19jau2fR13+ZlGWjIhvGY8dKVtMRAsqNRAG4ConovFvFtg+i/Oi6pHViljSAxm9w8=
trusted comment: polyvoice model signing | powerset_int8.onnx
2uP5mp2QmktBQ9Ef6miZpxT03EvqajAehheme+PFMU7Rj+PQ9uTE9HoI7HxOwR8k2uQuHNXYPOqn3cqWSdCyCg==
'''

[models.cam_pp_int8]
url      = "https://github.com/ekhodzitsky/polyvoice/releases/download/v0.6.0-alpha.2/cam_pp_int8.onnx"
sha256   = "cca48a4b36c1b46e48432b1eb1461dd69f9cf113cf506f3f660de808c93b9a85"
size     = 8803007
filename = "cam_pp_int8.onnx"
calibration = "voxconverse_dev_500_samples_seed_42"
license  = "Apache-2.0"
license_url = "https://huggingface.co/Wespeaker/wespeaker-voxceleb-campplus"
provenance = "INT8 quant of Wespeaker CAM++"
adapter_type = "cam++"
version  = "1.0-int8"
sample_rate = 16000
embedding_dim = 512
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhRQLqEqyoP/TNkcz2seLIK19JzqgIbPWHKFDVHMTA4+2hdmwZA5t0M6msDTE8LPEQXVpmqlb+jC4IzfAm518JgE=
trusted comment: polyvoice v0.6.0-alpha.3 | cam_pp_int8.onnx
XHIN9IgSVAhMFVb0TazR7NjXwO/ba2iCHQzrXsJEJ6hM4KpnrpjUgFaxTHkH4nVxlq8/5Pp4ODUc4qku3cKDAA==
'''

[models.resnet34_int8]
url      = "https://github.com/ekhodzitsky/polyvoice/releases/download/models-int8-v2/resnet34_int8.onnx"
sha256   = "24b58559fefb2af624a5d371c43ebae891a9a8ca363b2f9e7c31fd8e440a36b3"
size     = 6765058
filename = "resnet34_int8.onnx"
calibration = "voxconverse_dev_fbank_cmvn_rttm_segments_1819"
license  = "Apache-2.0"
license_url = "https://huggingface.co/Wespeaker/wespeaker-voxceleb-resnet34"
provenance = "INT8 quant of Wespeaker ResNet34"
adapter_type = "wespeaker-resnet34"
version  = "1.0-int8"
sample_rate = 16000
embedding_dim = 256
signature = '''
untrusted comment: signature from minisign secret key
RUQGu9FvZMmIhbyEbyzNuaI/cMPl2FynsjuPIz8pYuz56z6foEDqfdGOk4aMvB3pecxJQ4T16PmSK55buKIsMRDdrudPy498CwI=
trusted comment: polyvoice model signing | resnet34_int8.onnx
vthsMhwnQ8gSFSnTRWU9j3JiuQAQpAHsf7Tc1vKTad0o/IjuuPu7YPXY3vh/TvwaPMDNm3FqHgcKZ98VoMn6AA==
'''

# Optional ERes2NetV2 (Apache-2.0, ungated). Short-utterance-oriented 192-d
# embedder. NEVER profile-default / never bundled. SHA-256 from upstream LFS oid.
# Signature omitted until a release-signed asset exists.
[models.eres2netv2]
url      = "https://huggingface.co/csukuangfj/speaker-embedding-models/resolve/main/3dspeaker_speech_eres2netv2_sv_zh-cn_16k-common.onnx?download=true"
sha256   = "bf1a75b9930474cf3389ef415e6e5d38ca96fea4a3a00f7e301d080a58ee2239"
size     = 71441526
filename = "3dspeaker_speech_eres2netv2_sv_zh-cn_16k-common.onnx"
license  = "Apache-2.0"
license_url = "https://www.apache.org/licenses/LICENSE-2.0"
provenance = "3D-Speaker / ModelScope iic speech_eres2netv2_sv_zh-cn_16k-common; ONNX mirror csukuangfj/speaker-embedding-models"
adapter_type = "eres2netv2"
version  = "1.0"
sample_rate = 16000
embedding_dim = 192

# Optional CAM++ trained on 200k Chinese speakers (domain variant for CJK audio).
# Same fbank pipeline as default CAM++; select via adapter "cam++-zh".
[models.cam_pp_zh]
url      = "https://huggingface.co/csukuangfj/speaker-embedding-models/resolve/main/3dspeaker_speech_campplus_sv_zh-cn_16k-common.onnx?download=true"
sha256   = "f682b514c05d947ee3fa91cd6ec6c5c7543479a128373fa29b1faedccd21fd11"
size     = 28281138
filename = "3dspeaker_speech_campplus_sv_zh-cn_16k-common.onnx"
license  = "Apache-2.0"
license_url = "https://www.apache.org/licenses/LICENSE-2.0"
provenance = "3D-Speaker / ModelScope campplus zh-cn common; ONNX mirror csukuangfj/speaker-embedding-models"
adapter_type = "cam++-zh"
version  = "1.0"
sample_rate = 16000
embedding_dim = 192

# Optional E2E Streaming Sortformer v2 (≤4 speakers). NEVER default, NEVER
# bundled, NEVER pulled by profile resolution or default CI. Download only via
# explicit ModelRegistry::ensure("sortformer_v2") under the `sortformer` feature.
# Signature omitted until a release artifact is signed with the project key —
# SHA-256 alone gates integrity for this optional download path.
[models.sortformer_v2]
url      = "https://huggingface.co/cgus/diar_streaming_sortformer_4spk-v2-onnx/resolve/main/diar_streaming_sortformer_4spk-v2.onnx?download=true"
sha256   = "7dbfc7cba4615e07b679f7d65b5e0edd22a4b7b1ab69505a594f2f90421bd9c1"
size     = 492242946
filename = "diar_streaming_sortformer_4spk-v2.onnx"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "NVIDIA diar_streaming_sortformer_4spk-v2 (CC-BY-4.0); community ONNX export by Altunenes for parakeet-rs, mirrored at cgus/diar_streaming_sortformer_4spk-v2-onnx"
adapter_type = "sortformer-v2"
version  = "2.0"
sample_rate = 16000
num_speakers = 4

# ---------------------------------------------------------------------------
# VBx PLDA weights (six precomputed .npy files, ~265 KB total). NOT profile
# defaults: resolved only when the `vbx` clusterer is selected and no local
# `--vbx-plda-dir` / POLYVOICE_VBX_PLDA_DIR is set. Signature omitted until a
# release engineer signs with the project minisign key — SHA-256 alone gates
# integrity for this optional download path (same pattern as sortformer_v2 /
# eres2netv2). Hosted from the commit that introduced the fixtures; pin keeps
# the URL content-stable even if master moves.
# ---------------------------------------------------------------------------
[models.vbx_plda_transform]
url      = "https://raw.githubusercontent.com/ekhodzitsky/polyvoice/0623ae0b6773db7731f0fb3e75d8950347be94db/fixtures/vbx-plda/plda_transform.npy"
sha256   = "90261469714415743f4b8a86ee6b89466db858bde3c5944367cccfb7abd34f14"
size     = 131200
filename = "plda_transform.npy"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "pyannote speaker-diarization-community-1 PLDA params, diagonalized offline via scripts/build-vbx-plda.py; fixtures/vbx-plda"
adapter_type = "vbx-plda"
version  = "1.0"

[models.vbx_plda_phi_computed]
url      = "https://raw.githubusercontent.com/ekhodzitsky/polyvoice/0623ae0b6773db7731f0fb3e75d8950347be94db/fixtures/vbx-plda/plda_phi_computed.npy"
sha256   = "6ef7cf2f5a23a45b66f440f9a996a4cf5c047b369829af695d50ef18aa0a35e3"
size     = 1152
filename = "plda_phi_computed.npy"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "pyannote speaker-diarization-community-1 PLDA params, diagonalized offline via scripts/build-vbx-plda.py; fixtures/vbx-plda"
adapter_type = "vbx-plda"
version  = "1.0"

[models.vbx_plda_mean1]
url      = "https://raw.githubusercontent.com/ekhodzitsky/polyvoice/0623ae0b6773db7731f0fb3e75d8950347be94db/fixtures/vbx-plda/plda_mean1.npy"
sha256   = "e424c0c352182aa8e0f555dec1f3b30e29a20b9ed6b25d339f112af92e51e36f"
size     = 2176
filename = "plda_mean1.npy"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "pyannote speaker-diarization-community-1 PLDA params, diagonalized offline via scripts/build-vbx-plda.py; fixtures/vbx-plda"
adapter_type = "vbx-plda"
version  = "1.0"

[models.vbx_plda_mean2]
url      = "https://raw.githubusercontent.com/ekhodzitsky/polyvoice/0623ae0b6773db7731f0fb3e75d8950347be94db/fixtures/vbx-plda/plda_mean2.npy"
sha256   = "6f6fb708a2037197b5b84ffeaa8f140cb878088fbecd6ab042ad26a7691bd2cf"
size     = 640
filename = "plda_mean2.npy"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "pyannote speaker-diarization-community-1 PLDA params, diagonalized offline via scripts/build-vbx-plda.py; fixtures/vbx-plda"
adapter_type = "vbx-plda"
version  = "1.0"

[models.vbx_plda_lda]
url      = "https://raw.githubusercontent.com/ekhodzitsky/polyvoice/0623ae0b6773db7731f0fb3e75d8950347be94db/fixtures/vbx-plda/plda_lda.npy"
sha256   = "e20c9b012bebd1aabda5a38a127e63a43cf35debdc502715fc143e2fb6bc3c4b"
size     = 131200
filename = "plda_lda.npy"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "pyannote speaker-diarization-community-1 PLDA params, diagonalized offline via scripts/build-vbx-plda.py; fixtures/vbx-plda"
adapter_type = "vbx-plda"
version  = "1.0"

[models.vbx_plda_mu]
url      = "https://raw.githubusercontent.com/ekhodzitsky/polyvoice/0623ae0b6773db7731f0fb3e75d8950347be94db/fixtures/vbx-plda/plda_mu.npy"
sha256   = "d286d48acf99bbc1ed1502fed0a3e361ae5626ce1870c8be9f7397c5e47886c6"
size     = 1152
filename = "plda_mu.npy"
license  = "CC-BY-4.0"
license_url = "https://creativecommons.org/licenses/by/4.0/"
provenance = "pyannote speaker-diarization-community-1 PLDA params, diagonalized offline via scripts/build-vbx-plda.py; fixtures/vbx-plda"
adapter_type = "vbx-plda"
version  = "1.0"