feat(diarizer): Nemotron 3 Diarization support (8-speaker streaming Sortformer)

[Alex-Wengg]

Summary

Swift runtime for NVIDIA's Nemotron 3 Diarization (early-access preview on HF): 8 speakers, arrival-order channels, 10 ms output resolution, streaming and offline profiles. Code-only — model weights are not included; they load from a local directory and move to HuggingFace auto-download once NVIDIA's public release lands (converted models are staged and ready).

What's included

• Nemotron3Diarizer / Nemotron3StateUpdater — port of NeMo streaming_update_async at batch 1: fixed-capacity speaker-cache/FIFO state, score-based cache compression using the checkpoint's learned silence embedding, first-vs-later compression prediction freezing, NeMo-exact tail chunking, 10 ms high-resolution output extracted before state mutation. Closed-loop output verified against the NeMo/PyTorch reference (99.995% frame agreement on a 120 s fixture; remaining deltas are file-tail padding alignment).
• Nemotron3Models — local-dir CoreML loading; stride-aware vDSP_mmov output readback (model outputs are fp16 with padded rows — naive reads either scramble or run ~40x slower); per-chunk autoreleasepool (long ANE-route runs otherwise exhaust the IOSurface pool — same failure class as Diarization benchmark reports 80.8% DER on AMI ES2004a (target <30%) #752); optional split-graph mode (feature stacking + the 1024→512 projection run host-side via one cblas_sgemm) yielding a 100% ANE-resident pure-fp transformer graph that avoids an ANECCompile limit on long chunk inputs.
• Presets — the four model-card profiles (offline/low/verylow/ultra) plus chunk-ladder profiles (fast, fast24, fast32, fast128, efficient) and -int8 / -split selectors. Optional VAD gating (speechMask) skips inference across silence for sparse audio while preserving the output timeline.
• CLI — nemotron3-diarize, nemotron3-benchmark (AMI/VoxConverse harness, compute-unit routing, config sweep flags), nemotron3-batch (concurrent GPU workers).
• Tests — state-updater semantics (FIFO pop/flush, cache compression, silence slots), feature-loader tail handling, and padded-stride tensor-layout regression tests.

Notes

• Benchmark figures are intentionally deferred to post-release documentation per the model's early-access evaluation terms; accuracy and throughput tables will land with the weights.
• The mel frontend reuses the existing AudioMelSpectrogram (same 128-mel / 10 ms family as Nemotron ASR).

Testing

• swift build -c release clean; swift-format lint clean.
• Unit tests cover the state updater, feature loader, and tensor-layout regressions (CI).
• End-to-end parity against the NeMo reference validated on real audio via the benchmark harness.

[github-actions]

Supertonic3 Smoke Test ✅

Check	Result
Build	✅
Model download (incl. VectorEstimatorVariants/ int4 buckets)	✅
Model load	✅
Synthesis pipeline (--ve-variant int4)	✅
Output WAV	✅ (364.7 KB)

_{Runtime: 0m27s}

_{Note: CI VMs lack a physical Neural Engine; the ANE-bucketed VectorEstimator falls back to CPU here. This validates download + variant resolution + synthesis, not ANE residency/perf.}

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (153.8 KB)

_{Runtime: 0m6s}

_{Note: PocketTTS uses CoreML MLState (macOS 15) KV cache + Mimi streaming state. CI VM lacks physical GPU — audio quality and performance may differ from Apple Silicon.}

[github-actions]

Parakeet EOU Benchmark Results ✅

Status: Benchmark passed
Chunk Size: 320ms
Files Tested: 100/100

Performance Metrics

Metric	Value	Description
WER (Avg)	7.03%	Average Word Error Rate
WER (Med)	4.17%	Median Word Error Rate
RTFx	8.91x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	63.5s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.064s	Average chunk processing time
Max Chunk Time	0.127s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m14s • 08/30/2026, 05:36 PM EST}

_{RTFx = Real-Time Factor (higher is better) • Processing includes: Model inference, audio preprocessing, state management, and file I/O}

[github-actions]

Speaker Diarization Benchmark Results

Speaker Diarization Performance

Evaluating "who spoke when" detection accuracy

Metric	Value	Target	Status	Description
DER	15.1%	<30%	✅	Diarization Error Rate (lower is better)
JER	24.9%	<25%	✅	Jaccard Error Rate
RTFx	19.45x	>1.0x	✅	Real-Time Factor (higher is faster)

Diarization Pipeline Timing Breakdown

Time spent in each stage of speaker diarization

Stage	Time (s)	%	Description
Model Download	12.461	23.1	Fetching diarization models
Model Compile	5.341	9.9	CoreML compilation
Audio Load	0.106	0.2	Loading audio file
Segmentation	16.176	30.0	Detecting speech regions
Embedding	26.960	50.0	Extracting speaker voices
Clustering	10.784	20.0	Grouping same speakers
Total	53.950	100	Full pipeline

Speaker Diarization Research Comparison

Research baselines typically achieve 18-30% DER on standard datasets

Method	DER	Notes
FluidAudio	15.1%	On-device CoreML
Research baseline	18-30%	Standard dataset performance

Note: RTFx shown above is from GitHub Actions runner. On Apple Silicon with ANE:

• M2 MacBook Air (2022): Runs at 150 RTFx real-time
• Performance scales with Apple Neural Engine capabilities

_{🎯 Speaker Diarization Test • AMI Corpus ES2004a • 1049.0s meeting audio • 53.9s diarization time • Test runtime: 2m 43s • 08/30/2026, 05:37 PM EST}

[github-actions]

Offline VBx Pipeline Results

Speaker Diarization Performance (VBx Batch Mode)

Optimal clustering with Hungarian algorithm for maximum accuracy

Metric	Value	Target	Status	Description
DER	10.4%	<20%	✅	Diarization Error Rate (lower is better)
RTFx	10.00x	>1.0x	✅	Real-Time Factor (higher is faster)

Offline VBx Pipeline Timing Breakdown

Time spent in each stage of batch diarization

Stage	Time (s)	%	Description
Model Download	19.964	19.0	Fetching diarization models
Model Compile	8.556	8.2	CoreML compilation
Audio Load	0.072	0.1	Loading audio file
Segmentation	25.867	24.7	VAD + speech detection
Embedding	104.580	99.7	Speaker embedding extraction
Clustering (VBx)	0.113	0.1	Hungarian algorithm + VBx clustering
Total	104.885	100	Full VBx pipeline

Speaker Diarization Research Comparison

Offline VBx achieves competitive accuracy with batch processing

Method	DER	Mode	Description
FluidAudio (Offline)	10.4%	VBx Batch	On-device CoreML with optimal clustering
FluidAudio (Streaming)	17.7%	Chunk-based	First-occurrence speaker mapping
Research baseline	18-30%	Various	Standard dataset performance

Pipeline Details:

• Mode: Offline VBx with Hungarian algorithm for optimal speaker-to-cluster assignment
• Segmentation: VAD-based voice activity detection
• Embeddings: WeSpeaker-compatible speaker embeddings
• Clustering: PowerSet with VBx refinement
• Accuracy: Higher than streaming due to optimal post-hoc mapping

_{🎯 Offline VBx Test • AMI Corpus ES2004a • 1049.0s meeting audio • 130.6s processing • Test runtime: 2m 18s • 08/30/2026, 05:37 PM EST}

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	546.2x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	526.8x faster	50

Dataset Details

• MUSAN: Music, Speech, and Noise dataset - standard VAD evaluation
• VOiCES: Voices Obscured in Complex Environmental Settings - tests robustness in real-world conditions

✅: Average F1-Score above 70%

[github-actions]

Sortformer High-Latency Benchmark Results

ES2004a Performance (30.4s latency config)

Metric	Value	Target	Status
DER	30.3%	<35%	✅
Miss Rate	28.2%	-	-
False Alarm	0.9%	-	-
Speaker Error	1.2%	-	-
RTFx	14.5x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 3m 9s • 2026-08-30T21:44:02.825Z}

[github-actions]

ASR Benchmark Results ✅

Status: All benchmarks passed

Parakeet v3 (multilingual)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.57%	0.00%	6.68x	✅
test-other	1.19%	0.00%	3.77x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	6.19x	✅
test-other	1.00%	0.00%	3.50x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.69x	Streaming real-time factor
Avg Chunk Time	1.295s	Average time to process each chunk
Max Chunk Time	1.397s	Maximum chunk processing time
First Token	1.552s	Latency to first transcription token
Total Chunks	31	Number of chunks processed

Streaming (v2)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.70x	Streaming real-time factor
Avg Chunk Time	1.298s	Average time to process each chunk
Max Chunk Time	1.398s	Maximum chunk processing time
First Token	1.283s	Latency to first transcription token
Total Chunks	31	Number of chunks processed

_{Streaming tests use 5 files with 0.5s chunks to simulate real-time audio streaming}

_{25 files per dataset • Test runtime: 8m21s • 08/30/2026, 05:50 PM EST}

_{RTFx = Real-Time Factor (higher is better) • Calculated as: Total audio duration ÷ Total processing time
Processing time includes: Model inference on Apple Neural Engine, audio preprocessing, state resets between files, token-to-text conversion, and file I/O
Example: RTFx of 2.0x means 10 seconds of audio processed in 5 seconds (2x faster than real-time)}

Expected RTFx Performance on Physical M1 Hardware:

• M1 Mac: ~28x (clean), ~25x (other)
• CI shows ~0.5-3x due to virtualization limitations

_{Testing methodology follows HuggingFace Open ASR Leaderboard}