feat(diarizer/nemotron3): load monolithic presets from monolithic/v2 (M3 ANE compile fix)

[Alex-Wengg]

Refs #951.

The shipped monolithic Nemotron 3 bundles never compile for the ANE on M3 (h15g): ANECCompile() fails with InvalidMILProgram on every load, and CoreML falls back to a mixed CPU/ANE plan. On H15 the input-end packing (int16 gather, integer index casts) is split off to the CPU, so the first ANE function takes fp32 inputs, which h15g rejects.

• Monolithic presets now load from monolithic/v2/ on the HF repo: all five presets re-exported with fp16 inputs/outputs and gather-free state packing (mobius Need cocoapods support #100). Same checkpoint and weights.
• The originals stay in monolithic/, so older releases keep working and rollback is a one-line path change.
• No weightsVersion bump: the new path is a fresh download; split bundles and .bin assets stay cached.
• Unit test pins the subdirectory per preset.

Verification

• M3 Max (reporter, comment): fast32 v2 compiles, 100% frame decisions vs shipped, no ~3 s failed-compile retry per load.
• M5 Pro, AMI 16 meetings, shipped → v2: DER identical on .all for all five presets, within ±0.06 on ANE; RTFx within a few percent.
• Fresh download of v2 fast32 is byte-identical to the local build, --dump-preds exact; split presets unaffected.

🤖 Generated with Claude Code

[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: 0m29s}

_{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]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	730.6x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	723.7x 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]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (161.3 KB)

_{Runtime: 0m27s}

_{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	7.58x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	65.5s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.066s	Average chunk processing time
Max Chunk Time	0.131s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 2m11s • 09/24/2026, 08:09 PM EST}

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

[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%	3.96x	✅
test-other	1.19%	0.00%	3.10x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	4.48x	✅
test-other	1.00%	0.00%	3.21x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.55x	Streaming real-time factor
Avg Chunk Time	1.589s	Average time to process each chunk
Max Chunk Time	1.848s	Maximum chunk processing time
First Token	1.883s	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	0.58x	Streaming real-time factor
Avg Chunk Time	1.554s	Average time to process each chunk
Max Chunk Time	1.734s	Maximum chunk processing time
First Token	1.547s	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: 7m41s • 09/24/2026, 08:16 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}

[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	9.19x	>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	22.209	19.5	Fetching diarization models
Model Compile	9.518	8.3	CoreML compilation
Audio Load	0.094	0.1	Loading audio file
Segmentation	30.271	26.5	VAD + speech detection
Embedding	113.757	99.7	Speaker embedding extraction
Clustering (VBx)	0.154	0.1	Hungarian algorithm + VBx clustering
Total	114.156	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 • 144.2s processing • Test runtime: 2m 36s • 09/24/2026, 08:17 PM EST}

[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	22.02x	>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	13.543	28.4	Fetching diarization models
Model Compile	5.804	12.2	CoreML compilation
Audio Load	0.074	0.2	Loading audio file
Segmentation	14.285	30.0	Detecting speech regions
Embedding	23.808	50.0	Extracting speaker voices
Clustering	9.523	20.0	Grouping same speakers
Total	47.657	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 • 47.6s diarization time • Test runtime: 3m 23s • 09/24/2026, 08:19 PM EST}

[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	15.2x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 3m 36s • 2026-09-25T00:19:29.300Z}