fix(build): restore macOS x86_64 compile by removing unguarded Float16 (#875)

[Alex-Wengg]

Fixes #875.

Root cause

v0.15.6 added three files that address fp16 MLMultiArray storage via Swift.Float16, which is unavailable on macOS x86_64:

• ASR/Shared/LogitsArgmax.swift (shared by SenseVoice CTC + Paraformer)
• ASR/Paraformer/ParaformerManager.swift (rows(of:))
• TTS/NeuTts/NeuTtsSynthesizer.swift (seedState, floatBuffer)

This broke Intel swift build and downstream universal (arm64 x86_64) archives even for apps not using those features, since the sources are part of the monolithic FluidAudio target. Everything else in the tree was already portable via the established pattern (UInt16-backed pointers + vImage Planar16F conversions, as in KokoroAneArrays and ANEMemoryUtils); these files skipped it.

Changes

• New shared Float16Conversion enum (Shared/Float16Conversion.swift) wrapping vImageConvert_Planar16FtoPlanarF / vImageConvert_PlanarFtoPlanar16F; the three files now address half-precision storage as UInt16 bit patterns. No arch guards needed, bit-identical results, and the NeuTTS scalar per-element loops become vectorized passes as a side effect.
• Unit tests for the helper using hard-coded binary16 bit patterns, so the test file itself compiles on x86_64.
• Per the issue's suggestion, a new CI job in tests.yml: swift build --arch x86_64 cross-compile on the arm64 runner, so this regression class fails PR CI instead of surfacing on Intel machines after release.

Verification

• swift build --arch x86_64 on unpatched main reproduces every error from the issue verbatim (Paraformer 477/481, LogitsArgmax 35/44, NeuTTS 170/172/230/231).
• With this change, both swift build --arch x86_64 and native arm64 swift build complete cleanly.
• swift format lint clean on changed files.

[github-actions]

PocketTTS Smoke Test ✅

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

_{Runtime: 0m29s}

_{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	9.50x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	55.9s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.056s	Average chunk processing time
Max Chunk Time	0.112s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m39s • 08/22/2026, 07:57 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	26.23x	>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	10.552	26.4	Fetching diarization models
Model Compile	4.522	11.3	CoreML compilation
Audio Load	0.043	0.1	Loading audio file
Segmentation	11.999	30.0	Detecting speech regions
Embedding	19.998	50.0	Extracting speaker voices
Clustering	7.999	20.0	Grouping same speakers
Total	40.005	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 • 40.0s diarization time • Test runtime: 2m 50s • 08/22/2026, 07:58 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	12.66x	>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	19.318	23.3	Fetching diarization models
Model Compile	8.279	10.0	CoreML compilation
Audio Load	0.042	0.1	Loading audio file
Segmentation	23.425	28.3	VAD + speech detection
Embedding	82.706	99.7	Speaker embedding extraction
Clustering (VBx)	0.103	0.1	Hungarian algorithm + VBx clustering
Total	82.918	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 • 106.2s processing • Test runtime: 1m 56s • 08/22/2026, 07:58 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%	622.8x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	584.5x 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%	5.70x	✅
test-other	1.80%	0.00%	3.74x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	6.33x	✅
test-other	1.00%	0.00%	3.83x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.68x	Streaming real-time factor
Avg Chunk Time	1.340s	Average time to process each chunk
Max Chunk Time	1.466s	Maximum chunk processing time
First Token	1.582s	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	0.67x	Streaming real-time factor
Avg Chunk Time	1.345s	Average time to process each chunk
Max Chunk Time	1.481s	Maximum chunk processing time
First Token	1.334s	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: 9m6s • 08/22/2026, 08:04 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]

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: 0m58s}

_{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]

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	20.0x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 2m 45s • 2026-08-23T00:06:20.853Z}