File size: 8,287 Bytes
e919ff6 | 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 | {
"model": "Nemotron-3-Diarization (streaming Sortformer, 8 speakers)",
"license": "openmdw-1.1",
"source": {
"repo": "nvidia/Nemotron-3-Diarization",
"revision": "f667ed73aee57d40cc39428eb768b4fd87a0a29e",
"file": "model.safetensors",
"sha256": "c074d86335b3b794f8fa5edc25594558f128bdb3914d27806a3a5a2e44963cb6",
"bytes": 396954592,
"reference": "transformers Nemotron3DiarizationForAudioFrameClassification (fp32)"
},
"toolchain": {
"coreai-torch": "0.4.1",
"coreai-core": "1.0.0b2",
"torch": "2.9.0",
"macos": "27.0 (26A428)"
},
"graph": {
"function": "main",
"inputs": {
"packed": "[1, T, 512] float32: [speaker cache | FIFO | chunk rows (+ look-ahead)] left-packed at rows [0, L); rows [L, T) are ignored (zero them)",
"valid": "[1, T] float32: 1.0 for rows < L, 0.0 after"
},
"outputs": {
"logits": "[1, T*8, 8] float32 speaker logits at 10 ms; rows [0, L*8) are real"
},
"T": {
"streaming": 541,
"offline": 684
},
"host_side": "sigmoid, the 8x average pool to encoder rate, the speaker cache (AOSC + FIFO), segmentation"
},
"host": {
"frame_s": 0.01,
"encoder_frame_s": 0.08,
"subsampling": 8,
"num_speakers": 8,
"hidden": 512,
"profiles": {
"streaming": {
"T": 541,
"fifo_length": 264,
"update_period": 222,
"modes": {
"low_latency": {
"chunk": 9,
"lookahead": 4
},
"very_low_latency": {
"chunk": 6,
"lookahead": 2
},
"ultra_low_latency": {
"chunk": 3,
"lookahead": 1
}
},
"chunking": "per chunk mel from its own audio slice: chunk 0 = audio[:((c+la)*8-1)*160+200] center=True; chunk k>=1 starts at 160*k*c*8-256 and holds (c+la)*8*160+400 samples, center=False; the first chunk that would run past the audio takes the rest, is the last (no look-ahead) and emits its true mel frame count"
},
"offline": {
"T": 684,
"fifo_length": 40,
"update_period": 300,
"chunk": 340,
"lookahead": 40,
"chunking": "whole-recording mel (center=True) -> embeds -> chunks of 340 rows + up to 40 look-ahead rows; output cut to the mel frame count"
}
},
"step": "rows = [cache | FIFO | chunk]; logits = graph(rows)[:L*8]; probs = avg_pool8(sigmoid(logits)); emit logits[(n_cache+n_fifo)*8 : +min(n_chunk*8, n_emit)]; then the cache update",
"speaker_cache": {
"length": 264,
"silence_frames_per_speaker": 1,
"prediction_score_threshold": 0.25,
"latest_frames_score_boost": 0.05,
"min_positive_scores": 16,
"strong_boost": {
"frames": 24,
"add": 1.3862943611198906
},
"weak_boost": {
"frames": 48,
"add": 0.6931471805599453
},
"update": "transformers Nemotron3DiarizationSpeakerCache.update/_compress @ 4b28d51, line by line (host_loop.py)",
"score_dtype": "float64 from the float32 probabilities (transformers: float32; see host_loop.py)",
"topk_ties": "lower index first",
"sentinel": "the silence row (row N: silence_embeds, zero probs)"
},
"segments": "transformers extract_speaker_dict: per speaker, runs of sigmoid(logits) > 0.5; overlaps kept"
},
"mel": {
"sample_rate": 16000,
"preemphasis": 0.9700000286102295,
"preemphasis_first_sample": "kept",
"n_fft": 512,
"win_length": 400,
"hop": 160,
"window": "hann_window_400.f32le, centered in 512 (zeros at [0,56) and [456,512))",
"power": "|rFFT|^2 (as sqrt(re^2+im^2)^2 in float32)",
"n_mels": 128,
"fmin": 0.0,
"fmax": 8000.0,
"norm": "slaney",
"log": "log(mel + 2^-24)",
"log_guard": 5.960464477539063e-08,
"normalize": null,
"center_pad": "256 zeros each side when center=True (first chunk / offline), none otherwise",
"valid_frames": "center: n // 160; not center: (n - 512) // 160 + 1",
"stacking": "8 frames -> 1024 (zero-pad the frame count to a multiple of 8 in the log-mel domain)"
},
"assets": {
"embedder_projection.f32le": {
"shape": [
512,
1024
],
"what": "model.audio_tower.embedder.projection.weight (Linear 1024->512, no bias); embeds = stacked @ W.T",
"sha256": "add02ccfa230d319e82b55a49e6e2b0ae3d01fb326621b10e04e5aa5dd03d651",
"dtype": "float32 little-endian, C order"
},
"silence_embeds.f32le": {
"shape": [
512
],
"what": "silence_embeds: the speaker cache's silence row (1 per speaker)",
"sha256": "d4417b3c0eabdf7c47032fac2b5b5a7ee83d819a6ddda8fd8eaf74e2b5cc4ac7",
"dtype": "float32 little-endian, C order"
},
"mel_filters_128x257.f32le": {
"shape": [
128,
257
],
"what": "librosa.filters.mel(sr=16000, n_fft=512, n_mels=128, fmin=0, fmax=8000, norm='slaney'); bit-identical to the transformers feature extractor's",
"sha256": "bce5ec5f194a5913f6508cee5a85512e7bad2352db8fc28f5c6ff75af8b09137",
"dtype": "float32 little-endian, C order"
},
"hann_window_400.f32le": {
"shape": [
400
],
"what": "torch.hann_window(400, periodic=False); bit-identical to the feature extractor's",
"sha256": "c427e2029118cf789649e5a4d439b6115d0dd0cbf95dcd22f65e3c848add8c5b",
"dtype": "float32 little-endian, C order"
}
},
"bundles": {
"n3d_streaming_float16.aimodel": {
"profile": "streaming",
"T": 541,
"dtype": "float16",
"target": "macOS 27, Apple silicon GPU (compiled at load)",
"compute_preference": "gpu",
"bytes": 197982173,
"mb": 198.0,
"tree_sha256": "52563c95a40394c0b587e1550495b12bbb424dfcbfb5983a857f38db237d38f0",
"contract": {
"function": "main",
"inputs": {
"packed": {
"shape": [
1,
541,
512
],
"dtype": "float32"
},
"valid": {
"shape": [
1,
541
],
"dtype": "float32"
}
},
"outputs": {
"logits": {
"shape": [
1,
4328,
8
],
"dtype": "float32"
}
},
"states": []
}
},
"n3d_streaming_float16.h18p.aimodelc": {
"profile": "streaming",
"T": 541,
"dtype": "float16",
"target": "iOS 27, iPhone 17 Pro GPU (compiled ahead of time, h18p)",
"compute_preference": "gpu",
"bytes": 198238645,
"mb": 198.2,
"tree_sha256": "1e2840464bc246ab7508e1d5e4bf1b6a29e2f55159ff8fabf65c609303ccdb2d",
"compiled_from": "n3d_streaming_float16.aimodel",
"contract": "that of n3d_streaming_float16.aimodel",
"aot": {
"command": "xcrun coreai-build compile n3d_streaming_float16.aimodel --output <dir> --platform iOS --architecture h18p --preferred-compute gpu --min-deployment-version 27.0",
"coreai_build": "3600.83.1",
"ane_regions": 0
}
},
"n3d_offline_float16.aimodel": {
"profile": "offline",
"T": 684,
"dtype": "float16",
"target": "macOS 27, Apple silicon GPU (compiled at load)",
"compute_preference": "gpu",
"bytes": 198018769,
"mb": 198.0,
"tree_sha256": "93c99942a3b311824c90c6ccbdc0042eb259331a12a60ae414040963d001210d",
"contract": {
"function": "main",
"inputs": {
"packed": {
"shape": [
1,
684,
512
],
"dtype": "float32"
},
"valid": {
"shape": [
1,
684
],
"dtype": "float32"
}
},
"outputs": {
"logits": {
"shape": [
1,
5472,
8
],
"dtype": "float32"
}
},
"states": []
}
},
"n3d_offline_float16.h18p.aimodelc": {
"profile": "offline",
"T": 684,
"dtype": "float16",
"target": "iOS 27, iPhone 17 Pro GPU (compiled ahead of time, h18p)",
"compute_preference": "gpu",
"bytes": 198275246,
"mb": 198.3,
"tree_sha256": "8494215fba0bddc0d3a6b99673664ce6826d6e47882bdfde744abd7506be10ff",
"compiled_from": "n3d_offline_float16.aimodel",
"contract": "that of n3d_offline_float16.aimodel",
"aot": {
"command": "xcrun coreai-build compile n3d_offline_float16.aimodel --output <dir> --platform iOS --architecture h18p --preferred-compute gpu --min-deployment-version 27.0",
"coreai_build": "3600.83.1",
"ane_regions": 0
}
}
},
"notes": [
"compute_preference: load with SpecializationOptions(preferredComputeUnitKind: .gpu); the .h18p.aimodelc bundles were also compiled with --preferred-compute gpu.",
"The float32 bundles are for parity only (CPU) and are not shipped: export_n3d.py --dtype float32 [--profile offline] rebuilds them."
]
}
|