Nemotron 3 diarization — Windows x64 Intel NPU FP16
This representation preserves the upstream FP16 weights and expands 62
RotaryEmbedding and 62 SkipLayerNormalization contrib nodes into standard
ONNX operators. It uses fixed input shapes required by the Windows ML NPU path.
The graph and external-data sidecar are pinned by SHA-256 in manifest.json.
On an Intel Core Ultra 5 325 running Windows build 26200, the native runtime
loaded the graph with CPU fallback disabled. The Goodmeet speaker evaluator
reported NPU execution for diarization and embedding on all 36 AMI SDM items.
ONNX Runtime profiling recorded one OpenVINO provider kernel and no CPU
provider nodes for each graph on a five-second sample. Audio preparation and
speaker processing outside ONNX Runtime still use CPU.
Against the current product model with CPU diarization fallback, micro DER was
33.44% versus 33.34%, and median diarization real-time factor was 0.0073 versus
0.0561. One item regressed by 2.50 percentage points of DER. Identity metrics
were unchanged. A hybrid graph retaining CPU RotaryEmbedding islands scored
33.45% micro DER and 0.0110 median RTF on the same series, so the fully defused
graph is the preferred Intel NPU artifact. These far-field AMI results do not
establish quality on other microphones or hardware.
On a separate 16-item AMI headset-0 holdout, micro DER was 36.41% on NPU versus 36.85% for the packaged graph's CPU fallback. Median diarization RTF was 0.0070 versus 0.0574, with no failed items and unchanged WeSpeaker EER. This is one headset channel on one Intel host; product-microphone and other Windows hardware conditions remain unmeasured.
Source: onnx-community/Nemotron-3-Diarization-ONNX at
353b6f8ad2cac3580e982d7fbdf0a010786b0406.