Nemotron 3 Diarization · CoreML INT8

Compiled CoreML export of NVIDIA Nemotron 3 Diarization for Apple devices. It emits frame-level activity probabilities for up to eight speakers. It does not transcribe speech or identify people.

Model

Property	Value
Parameters	99.2 million in the source model
Weights	INT8 block-32 linear weights; FP16 convolution and activations
Format	Two compiled `.mlmodelc` graphs, no `.mlpackage`
File size	About 103 MiB combined
Audio	16 kHz mono, 128 mel bands
Chunk	30.4 seconds of mel features; 10 ms output frames
Speakers	Up to eight
Minimum deployment target	iOS 18

Files

File	Size	Purpose
`Nemotron3PreEncoder.mlmodelc/`	About 0.6 MiB	Mel-feature projection
`Nemotron3Head.mlmodelc/`	About 102 MiB	Transformer and speaker-activity head
`learnable_silence.f32`	2 KiB	Speaker-cache silence embedding
`config.json`	About 2 KiB	Shapes, cache settings, source revision, and export checks
`run_chunk.py`	About 2 KiB	First-chunk command line example
`LICENSE`	About 3 KiB	OpenMDW 1.1 license
`NOTICE`	Under 1 KiB	Source and copyright notice

Performance

Measured on an Apple M5 Pro running macOS 26.6.2 with one 30.4-second chunk from each of three locally held speech recordings. Times are the median of three warm CPU-only runs of the compiled model stages. Mel extraction, file loading, speaker-cache updates, and segment postprocessing are excluded. The comparison is against the final unquantized checkpoint; this is a numerical parity check, not a diarization error rate benchmark.

Measure	CoreML INT8	Meaning
Median model-stage time	101.6 ms	About 300× faster than the chunk's audio duration on CPU; lower is better
Mean absolute probability difference	0.00046–0.00082	Close to the source checkpoint; lower is better
Speaker-activity decision disagreement at 0.25 threshold	0.03–0.06% of frame/channel pairs	Quantization changes few decisions; lower is better

The speech-swift host runtime scored 11.20% pooled DER on 132 internal benchmark recordings (12.82 hours) across eight languages, using a 0.25-second collar, overlap included, and automatic speaker count. All files completed. This measures the full Swift audio-to-segment pipeline, not the first-chunk example below. Other hosts must reproduce the source model's cache and FIFO update algorithm to preserve speaker labels across chunks.

Usage

Install coremltools and numpy on macOS. Supply a NumPy file containing preprocessed 128-band log-mel features with shape [1, 3040, 128], matching the source NeMo preprocessor. The command below processes the first chunk with empty speaker memory:

python run_chunk.py --format coreml --bundle . --mel chunk.npy --output probabilities.npy

import coremltools as ct
import numpy as np

pre = ct.models.CompiledMLModel("Nemotron3PreEncoder.mlmodelc")
head = ct.models.CompiledMLModel("Nemotron3Head.mlmodelc")
# mel: float32 NumPy array [1, 3040, 128]; first chunk only
embeddings = pre.predict({"chunk": mel})["chunk_embeddings"]
packed = np.zeros((1, 684, 512), dtype=np.float32)
packed[:, :380] = embeddings
probabilities = head.predict({
    "packed_embeddings": packed,
    "packed_length": np.array([380], dtype=np.int32),
})["speaker_probabilities_10ms"][:, :3040]

Binarize speaker activity with 0.5 onset and offset thresholds, as in the source model's default postprocessor. The pred_score_threshold: 0.25 value in config.json is used to select speaker-cache frames; it is not the segment activity threshold. For later chunks, preserve the arrival-order speaker cache and FIFO state as described by the source model.

Source

Converted from nvidia/Nemotron-3-Diarization, revision a435e9867d79e789e90053f9b6d6834053af564a. NVIDIA's model description explains the architecture and benchmark protocol. The source model and these derived weights are distributed under OpenMDW 1.1; see LICENSE and NOTICE.

aufklarer/Nemotron-3-Diarization-100M-CoreML-INT8