fix(itn): link the bundled NeMo engine directly instead of dlopen(nil) discovery

[Alex-Wengg]

Fixes #839

Problem

TextNormalizer resolved the nemo_* symbols at runtime via dlopen(nil) + dlsym — a leftover from when the native ITN library was an optional, consumer-provided link. Since #790 the NemoTextProcessing xcframework ships with the package as a binary target, so runtime discovery is pure downside:

• Releases ≤ 0.15.5 shipped no artifact at all, making normalize() a silent no-op for every SwiftPM consumer (TextNormalizer is a silent no-op for SwiftPM consumers: no native ITN library ships, dlopen(nil) finds nothing, normalize() returns input unchanged (0.15.5) #839). v0.15.6 bundles the xcframework, which fixes standard builds.
• Even with the bundled library, dlsym silently regresses to the no-op in any app that strips global symbols from its executable (common in Release configurations) — the symbols are linked in but no longer discoverable.
• normalizeSentence(_:maxSpanTokens:) looked up nemo_normalize_sentence_with_max_span, which does not exist in the shipped v0.3.0 library, so maxSpanTokens was silently ignored.

Fix

Import CNemoTextProcessing and call the C API directly, exactly like the TTS-side NemoTextNormalizer already does. Availability becomes a link-time guarantee, so the silent-unavailability failure mode is impossible by construction rather than merely logged (issue ask #1).

• isNativeAvailable / isTnAvailable are kept (always true) for source compatibility with ≤ 0.15.6 consumers that branch on them.
• The max-span variant now calls nemo_normalize_sentence_with_options, so maxSpanTokens is actually honored.
• All the function-pointer plumbing is deleted (498 → ~260 lines); NLTagger sentinel masking is unchanged.
• Docs: removed the obsolete "link the library yourself" section — the engine works out of the box.

Tests

Tests that guarded on !isNativeAvailable became permanently-skipped dead code once the library is always linked. They now assert real engine output, including the exact issue repro:

XCTAssertEqual(normalizer.normalize("twelve dollars"), "$12")

Every expectation was verified against the bundled v0.3.0 engine (C driver against the static lib + a compiled Swift check through the real TextNormalizer): single-expression, sentence mode, max-span, ambiguous-word masking ("the period of growth" preserved), TN surface, and custom rules.

Note: consumers get this only with a release cut after merge — v0.15.6 already fixes standard builds, but a follow-up tag carries the strip-proof direct linking.

[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: 0m51s}

_{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	6.59x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	71.8s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.072s	Average chunk processing time
Max Chunk Time	0.144s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 2m20s • 08/19/2026, 10:05 AM EST}

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

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	360.0x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	396.7x 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]

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	14.8x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 5m 1s • 2026-08-19T14:09:02.933Z}

[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	18.28x	>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	14.484	25.2	Fetching diarization models
Model Compile	6.208	10.8	CoreML compilation
Audio Load	0.063	0.1	Loading audio file
Segmentation	17.216	30.0	Detecting speech regions
Embedding	28.693	50.0	Extracting speaker voices
Clustering	11.477	20.0	Grouping same speakers
Total	57.417	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 • 57.4s diarization time • Test runtime: 3m 15s • 08/19/2026, 10:13 AM EST}

[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%	3.38x	✅
test-other	1.19%	0.00%	3.18x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.34x	✅
test-other	1.40%	0.00%	3.22x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.54x	Streaming real-time factor
Avg Chunk Time	1.652s	Average time to process each chunk
Max Chunk Time	2.890s	Maximum chunk processing time
First Token	2.051s	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	0.59x	Streaming real-time factor
Avg Chunk Time	1.511s	Average time to process each chunk
Max Chunk Time	1.812s	Maximum chunk processing time
First Token	1.573s	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: 10m30s • 08/19/2026, 10:13 AM 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]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (168.8 KB)

_{Runtime: 0m32s}

_{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]

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	7.86x	>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	23.767	17.8	Fetching diarization models
Model Compile	10.186	7.6	CoreML compilation
Audio Load	0.083	0.1	Loading audio file
Segmentation	33.763	25.3	VAD + speech detection
Embedding	133.228	99.8	Speaker embedding extraction
Clustering (VBx)	0.127	0.1	Hungarian algorithm + VBx clustering
Total	133.551	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 • 167.1s processing • Test runtime: 2m 55s • 08/19/2026, 10:21 AM EST}