fix(diarizer/offline): warn on macOS 14 BNNS-crash-prone builds + document known OS issue (#878)

[Alex-Wengg]

Library-side portion of #878 (the #661/#817 BNNS crash family).

Context

@JulianPscheid's CI matrix in #878 proves the offline diarizer libBNNS crash (EXC_BAD_ACCESS in BNNSGraphContextExecute_v2 → _platform_memmove) is an Apple OS bug in the macOS 14 line:

• macos-14 (14.8.7): 1200/1200 crashes — every variant (unpatched v0.15.5, Core ML-owned FBank storage, fully serialized single-model runs), 211 matching .ips artifacts.
• macos-15 (15.7.7) and macos-26 (26.5.2): 0 crashes, same code, sha256-pinned audio. Comparison run twice.
• On VMs without an ANE every prediction takes the BNNS CPU path, so the crash is deterministic there; on Apple Silicon only ANE-fallback predictions hit BNNS, which is why it looks intermittent in the field.
• The serialization proposed in Concurrent SlidingWindowAsrManager + DiarizerManager prediction corrupts shared E5RT/BNNS state (EXC_BAD_ACCESS in libBNNS) on Apple Silicon #661 does not help: single-threaded, single-model runs still crash on macOS 14.

Changes

Since no library-side mitigation exists, the honest fix is diagnosis — same approach as #818 took for the (distinct) 26.4–26.5 Kokoro BNNS bug:

• OfflineDiarizerManager.initialize(models:) logs a warning on macOS 14 builds pointing at Offline diarizer BNNS crash (#661 family) reproduces deterministically on macOS 14; absent on 15 and 26 #878, so field crashes get attributed to the OS instead of usage patterns or FluidAudio versions. Predicate extracted as isBnnsCrashProneOS(_:onMacOS:) for testability. iOS is unflagged — no reproduction has been reported on the iOS 17 line.
• OfflineDiarizerOsAdvisoryTests — boundary coverage (13.6 ✗ / 14.0 ✓ / 14.8.7 ✓ / 15.0 ✗ / 26.5.2 ✗; iOS unflagged).
• Documentation/Diarization/GettingStarted.md — "Known OS issue" note in the Offline VBx section: macOS 14 can crash in BNNS regardless of usage pattern, fixed by updating to macOS 15+; explicitly distinguished from the Kokoro 26.4–26.5 class in KokoroAne.md.
• Cross-ref comment at the FBank .cpuOnly pin (OfflineDiarizerModels.swift), which is why the diarizer is the component that surfaces this on macOS 14.

Deliberately not changed: FBank routing on macOS 14. The matrix shows every compute-unit configuration crashes there, so a placement tweak would be an unverifiable non-fix; if we want to test whether .cpuAndGPU FBank reduces exposure on real hardware, the reporter's workflow_dispatch harness can take a revision input.

Verification

• swift build green, swift format lint clean on touched files.
• Advisory tests run in CI (XCTest unavailable locally).

Closes nothing on its own — #878 disposition (close as OS bug vs. keep open for Apple Feedback tracking) is up for discussion.

🤖 Generated with Claude Code

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (146.3 KB)

_{Runtime: 0m20s}

_{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]

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: 0m32s}

_{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]

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	8.98x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	55.5s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.056s	Average chunk processing time
Max Chunk Time	0.111s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m35s • 08/31/2026, 11:26 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	26.08x	>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	13.039	32.4	Fetching diarization models
Model Compile	5.588	13.9	CoreML compilation
Audio Load	0.055	0.1	Loading audio file
Segmentation	12.067	30.0	Detecting speech regions
Embedding	20.111	50.0	Extracting speaker voices
Clustering	8.045	20.0	Grouping same speakers
Total	40.234	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 • 40.2s diarization time • Test runtime: 3m 6s • 08/31/2026, 11:29 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	10.24x	>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	22.510	22.0	Fetching diarization models
Model Compile	9.647	9.4	CoreML compilation
Audio Load	0.084	0.1	Loading audio file
Segmentation	28.199	27.5	VAD + speech detection
Embedding	102.082	99.6	Speaker embedding extraction
Clustering (VBx)	0.145	0.1	Hungarian algorithm + VBx clustering
Total	102.477	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 • 130.4s processing • Test runtime: 2m 24s • 08/31/2026, 11:34 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	19.4x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 5m 16s • 2026-09-01T03:35:46.995Z}

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	554.8x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	490.3x 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]

ASR Benchmark Results ✅

Status: All benchmarks passed

Parakeet v3 (multilingual)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.57%	0.00%	4.31x	✅
test-other	1.19%	0.00%	3.05x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.16x	✅
test-other	1.00%	0.00%	3.10x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.58x	Streaming real-time factor
Avg Chunk Time	1.527s	Average time to process each chunk
Max Chunk Time	1.849s	Maximum chunk processing time
First Token	1.835s	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	0.52x	Streaming real-time factor
Avg Chunk Time	1.729s	Average time to process each chunk
Max Chunk Time	2.046s	Maximum chunk processing time
First Token	1.768s	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: 11m8s • 08/31/2026, 11:43 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}